WO2022262080A1 - 一种对话关系处理方法、计算机及可读存储介质 - Google Patents

一种对话关系处理方法、计算机及可读存储介质 Download PDF

Info

Publication number
WO2022262080A1
WO2022262080A1 PCT/CN2021/108503 CN2021108503W WO2022262080A1 WO 2022262080 A1 WO2022262080 A1 WO 2022262080A1 CN 2021108503 W CN2021108503 W CN 2021108503W WO 2022262080 A1 WO2022262080 A1 WO 2022262080A1
Authority
WO
WIPO (PCT)
Prior art keywords
sample
argument
text
dialogue
relationship
Prior art date
Application number
PCT/CN2021/108503
Other languages
English (en)
French (fr)
Inventor
赵田阳
闫昭
Original Assignee
腾讯云计算（北京）有限责任公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯云计算（北京）有限责任公司 filed Critical 腾讯云计算（北京）有限责任公司
Publication of WO2022262080A1 publication Critical patent/WO2022262080A1/zh
Priority to US18/132,723 priority Critical patent/US20230244878A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G06F40/35—Discourse or dialogue representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/274—Converting codes to words; Guess-ahead of partial word inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities

Definitions

  • the present application relates to the field of computer technology, and in particular to a dialog relationship processing method, a computer and a readable storage medium.
  • the process of data processing involves the analysis of arguments in conversational data.
  • the task of extracting argument relations is used to determine conversational data.
  • Argument relationships between arbitrary arguments in the data are used.
  • the dialogue text and argument pairs are generally extracted with semantic features to obtain relevant semantic feature representations, and then based on the semantic feature representations, the argument relations corresponding to the argument pairs are predicted.
  • Embodiments of the present application provide a dialog relationship processing method, a computer, and a readable storage medium, which can improve the prediction accuracy of argument relationships in a dialog scene.
  • an embodiment of the present application provides a dialog relationship processing method, the method is applied to a computer device, and the method includes:
  • the trigger word Predict the trigger word based on the semantic information of the sample text and the actual argument relationship, and determine the third loss based on the trigger word prediction result, the actual argument relationship is the labeled argument relationship corresponding to the sample argument, and the trigger word
  • the word prediction result is used to characterize the position of the trigger word in the described sample dialogue text;
  • the initial relationship prediction model is trained to obtain a dialogue relationship prediction model.
  • an embodiment of the present application provides a dialog relationship processing method, the method is applied to a computer device, and the method includes:
  • the dialogue relationship prediction model is obtained through training using the dialogue relationship processing method described in the above aspect.
  • an embodiment of the present application provides a dialog relationship processing device, which includes:
  • the model input module and the semantic information extraction module are used to extract the semantic features of the sample dialogue text and sample argument pairs through the initial relationship prediction model to obtain the semantic information of the sample text, and each sample argument in the sample argument pair belongs to the Describe sample dialogue text;
  • a relationship processing module and a first loss generation module configured to perform argument relationship prediction based on the semantic information of the sample text, and determine the first loss based on the argument relationship prediction result, the argument relationship prediction result being used to characterize the sample the relationship between arguments;
  • a character acquisition module, a phrase prediction module, and a second loss generation module configured to perform hidden character prediction based on the sample text semantic information, and determine a second loss based on the hidden character prediction result, the hidden character prediction result being used to characterize the Hidden characters in sample dialogue text;
  • a trigger detection generation module, a sequence tagging module, and a third loss generation module are used to perform trigger word prediction based on the semantic information of the sample text and actual argument relationships, and determine a third loss based on the trigger word prediction results, and the actual argument
  • the relationship is a labeled argument relationship corresponding to the sample argument, and the trigger word prediction result is used to characterize the position of the trigger word in the sample dialogue text;
  • a model training module configured to train the initial relationship prediction model based on the first loss, the second loss, and the third loss to obtain a dialogue relationship prediction model.
  • an embodiment of the present application provides a dialog relationship processing device, which includes:
  • the target model input module is used to input the target dialogue text and the target argument into the dialogue relationship prediction model
  • the target semantic acquisition module is used to extract the semantic features of the target dialogue text and the target argument pair in the dialogue relationship prediction model, and obtain the target text semantic information corresponding to the target dialogue text, and each target argument in the target argument pair Belongs to said target dialogue text;
  • the target relationship prediction module is used to predict the argument relationship of the target text semantic information based on the relationship prediction network in the dialogue relationship prediction model, and obtain the target argument relationship between the target arguments.
  • the dialogue relationship prediction model adopts the above aspects obtained through the training of the dialog relationship processing method.
  • an embodiment of the present application provides a computer device, including a processor, a memory, and an input and output interface;
  • the processor is connected to the memory and the input/output interface respectively, wherein the input/output interface is used to receive data and output data, the memory is used to store the computer program, and the processor is used to call the computer program, so that the computer device including the processor executes the program.
  • the dialog relationship processing method in one aspect of the application embodiment.
  • an embodiment of the present application provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and the computer program is adapted to be loaded and executed by a processor, so that a computer device having the processor executes The dialog relationship processing method in one aspect of the embodiment of the present application.
  • an embodiment of the present application provides a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the dialog relationship processing method provided in the various optional implementation manners above.
  • the initial phrase prediction network predicts the hidden speaker arguments in the sample dialogue text, and the trigger words
  • the prediction network predicts the trigger words that can guide the argument relationship in the sample dialogue text, so as to introduce additional second loss and third loss in the loss, and provide auxiliary information for the argument relationship prediction: speaker features and trigger word features, This enables the initial relationship prediction model to learn effective information that is more conducive to predicting the argument relationship, thereby improving the prediction accuracy of the argument relationship.
  • FIG. 1 is a network interaction architecture diagram of dialog relationship processing provided by an embodiment of the present application
  • FIG. 2 is a schematic diagram of a model training scenario provided by an embodiment of the present application.
  • Fig. 3 is a schematic structural diagram of a model provided by an embodiment of the present application.
  • FIG. 4 shows a flow chart of a dialog relationship processing method shown in an exemplary embodiment of the present application
  • Fig. 5 is a flow chart of a dialog relationship processing method provided by another exemplary embodiment of the present application.
  • FIG. 6 is a model architecture diagram of an initial relationship prediction model provided by an exemplary embodiment of the present application.
  • FIG. 7 is a schematic diagram of the principle of an initial relationship prediction model provided by an exemplary embodiment of the present application.
  • FIG. 8 is a flow chart of a method for processing dialog relationships provided by an exemplary embodiment of the present application.
  • Fig. 9 is a schematic diagram of an apparatus for processing dialogue relations provided by an exemplary embodiment of the present application.
  • Fig. 10 is a schematic diagram of another dialog relationship processing device provided by the embodiment of the present application.
  • FIG. 11 is a schematic structural diagram of a computer device provided by an embodiment of the present application.
  • FIG. 1 is a network interaction architecture diagram of dialog relationship processing provided in the embodiment of the present application.
  • the embodiment of the present application can be implemented by a computer device 101, which can interact with the user Data exchange is performed between devices (such as the user equipment 102a, the user equipment 102b, and the user equipment 102c, etc.).
  • the computer device 101 can obtain the sample dialogue text and the sample argument pairs in the sample dialogue text from each user equipment, based on the sample dialogue text and the sample argument pairs Realize the training of the model, and obtain the dialog relationship prediction model for predicting the argument pair relationship.
  • the sample dialogue text may be obtained from each user device, or may be obtained directly from the computer device 101 (that is, stored in the computer device 101), or may be obtained based on cloud storage technology, or may be Acquisition based on the blockchain network is not limited here.
  • the computer device 101 is a device for model prediction, that is, the computer device 101 runs a pre-trained dialog relationship prediction model, then the computer device 101 can respond to a relationship prediction request sent by any user device, for example, in response to the user
  • the relationship prediction request sent by the device 102a obtains the target dialogue text and the target argument pair included in the relationship prediction request, predicts the target dialogue text and the target argument pair based on the dialogue relationship prediction model, and obtains the target theory of the target argument pair. meta relationship.
  • the computer device 101 may also acquire the target dialogue text and the target argument pair stored in the computer device 101, predict the target dialogue text and the target argument pair based on the dialogue relationship prediction model, and obtain the target argument pair The target argument relationship of , etc.
  • the data involved in this application can be provided by the user equipment, or can be stored in the computer equipment 101, or can be stored based on cloud storage technology, or can be stored in the block chain network, not here Do limit.
  • the embodiment of the present application adds a speaker prediction task and a trigger word prediction task to the training task of the dialogue relationship prediction model, wherein the speaker prediction task is used to predict the sample dialogue The masked speaker arguments in the text, and the trigger word prediction task is used to predict the words representing the argument relationship in the sample dialogue text, both of which can assist the dialogue relationship prediction task.
  • FIG. 2 is a schematic diagram of a model training scenario provided by an embodiment of the present application.
  • the computer device can obtain a model training sample 201, the model training sample 201 includes a sample dialogue text and a sample argument pair in the sample dialogue text, and the sample dialogue text and the sample argument pair are input into the initial relationship prediction model, Extract the semantic features of the sample dialogue text and sample argument pairs in the initial relationship prediction model to obtain the sample text semantic information corresponding to the sample dialogue text, and input the sample text semantic information into the task processing area in the initial relationship prediction model, including relationship prediction Task area 2021, phrase prediction task area 2022, trigger word prediction task area 2023, etc.
  • the sample argument pair includes a first sample argument and a second sample argument.
  • the relationship prediction task area 2021 is used to perform relationship extraction on sample argument pairs, and the relationship extraction is a technology for identifying and determining specific relationships between entity pairs from natural language texts.
  • the argument relationship is predicted based on the semantic information of the sample text
  • the predicted argument relationship between the first sample argument and the second sample argument is obtained
  • the first sample argument The first loss is generated based on the predicted and actual argument relations between the arguments of the second sample and the actual argument relations.
  • the masked hidden state corresponding to the hidden character in the sample dialogue text is obtained from the semantic information of the sample text, and then the predicted character corresponding to the hidden character is predicted based on the masked hidden state, and then according to the hidden character and the predicted Characters generate a second loss.
  • the trigger word prediction task area 2023 the trigger word detection text data is generated according to the actual argument relationship and the sample text semantic information, and the prediction sequence annotation in the trigger word detection text data is predicted.
  • the prediction annotation sequence is used to indicate each The trigger word type to which the sample dialogue phrase belongs; the third loss is generated by detecting the actual sequence annotation and the predicted sequence annotation in the text data according to the trigger word.
  • the relationship prediction task can more comprehensively obtain the dialogue text (such as sample dialogue text or target dialogue text, etc.) information, thereby improving the prediction accuracy of this relationship prediction task.
  • the initial relationship prediction model includes an initial relationship prediction network, an initial phrase prediction network, and an initial trigger word prediction network.
  • the initial relationship prediction model also includes an initial language representation network.
  • the initial relationship prediction network, initial phrase The input of the prediction network and the initial trigger word prediction network is obtained based on the output of the initial language representation network.
  • a dialogue relationship prediction model is generated, wherein the model parameters use general variables to establish the relationship between functions and variables A quantity.
  • the model parameters are usually a matrix of real numbers.
  • the model parameters can also be data in other formats.
  • the dialog relationship prediction model may include a language representation network, a relationship prediction network, a phrase prediction network and a trigger word prediction network.
  • the dialog relationship prediction model may only include language representation network and relationship prediction network, etc., so as to simplify the model structure.
  • the phrase prediction network and the trigger word prediction network are used to supplement the characteristic information of the relationship prediction network, thereby improving the prediction accuracy of the relationship prediction network.
  • the relationship prediction task area refers to the initial relationship prediction network and the area where the relationship prediction network is located, that is to say, the network in the relationship prediction task area is called the initial relationship prediction network before training, and is called the relationship prediction network after training.
  • the phrase prediction task area refers to the area where the initial phrase prediction network and the phrase prediction network are located, that is, the phrase prediction
  • the network in the task area is called the initial phrase prediction network before training, and it is called the phrase prediction network after training, which is used to predict hidden phrases in the dialogue text
  • the trigger word prediction task area refers to the initial trigger word prediction network and the trigger word prediction network.
  • the area where the word prediction network is located, that is to say, the network in the trigger word prediction task area is called the initial trigger word prediction network before training, and is called the trigger word prediction network after training, which is used to predict whether each dialogue phrase in the dialogue text is are trigger words.
  • the computer equipment or user equipment mentioned in the embodiments of the present application includes but not limited to terminal equipment or servers.
  • a computer device or a user device may be a server or a terminal device, or a system composed of a server and a terminal device.
  • the terminal device mentioned above can be an electronic device, including but not limited to mobile phones, tablet computers, desktop computers, notebook computers, handheld computers, vehicle-mounted devices, augmented reality/virtual reality (Augmented Reality/Virtual Reality, AR /VR) equipment, head-mounted displays, smart TVs, wearable devices, smart speakers, digital cameras, cameras, and other mobile Internet devices (Mobile Internet Device, MID) with network access capabilities, or in scenarios such as trains, ships, and flights terminal equipment, etc.
  • augmented reality/virtual reality Augmented Reality/Virtual Reality, AR /VR
  • head-mounted displays smart TVs
  • wearable devices smart speakers
  • digital cameras cameras
  • cameras digital cameras
  • other mobile Internet devices Mobile Internet Device, MID
  • the server mentioned above can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, Cloud servers for basic cloud computing services such as network services, cloud communications, middleware services, domain name services, security services, vehicle-road collaboration, content delivery network (Content Delivery Network, CDN), and big data and artificial intelligence platforms.
  • cloud services cloud databases, cloud computing, cloud functions, cloud storage, Cloud servers for basic cloud computing services such as network services, cloud communications, middleware services, domain name services, security services, vehicle-road collaboration, content delivery network (Content Delivery Network, CDN), and big data and artificial intelligence platforms.
  • CDN Content Delivery Network
  • FIG. 4 shows a flow chart of a method for processing dialogue relations shown in an exemplary embodiment of the present application.
  • the embodiment of the present application takes the application of the method to the computer device shown in FIG. 1 as an example for illustration.
  • Methods include:
  • Step 401 Extract semantic features of the sample dialogue text and sample argument pairs through the initial relationship prediction model to obtain the semantic information of the sample text, and each sample argument in the sample argument pair belongs to the sample dialogue text.
  • the initial relationship prediction model includes an initial language representation network, an initial relationship prediction network, an initial phrase prediction network, and an initial trigger word prediction network, wherein the initial language representation network is used to extract sample text semantic information (that is, text semantic feature representation), The initial relationship prediction network, the initial phrase prediction network and the initial trigger word prediction network share the semantic information of the sample text.
  • the initial relationship prediction network predicts the argument relationship based on the semantic information of the sample text, and the initial phrase prediction network predicts the sample text based on the semantic information of the sample text.
  • Hidden characters in the dialogue text the initial trigger word prediction network predicts whether each sample phrase in the sample dialogue text is a trigger word based on the semantic information of the sample text.
  • the initial relationship prediction model also includes a semantic enhancement network, which is used to perform semantic enhancement on the extracted sample text semantic information.
  • the initial language representation network can use Bidirectional Encoder Representations from Transformers (BERT), autoregressive pre-training model (Embedding from Language Model, ELMo) or generative pre-training model (Generative Pre-Training, GPT/GPT2), etc., the embodiment of the present application does not limit the network type used by the initial language representation network.
  • BERT Bidirectional Encoder Representations from Transformers
  • ELMo autoregressive pre-training model
  • GPT/GPT2 Generic Pre-Training
  • the computer device inputs the sample dialogue text and sample argument pairs into the initial relationship prediction model, and the initial relationship prediction in the initial relationship prediction model
  • the network extracts the semantic features of the sample dialogue text and sample argument pairs to obtain the semantic information of the sample text.
  • the information is in the form of feature vectors.
  • the training sample set in the embodiment of the present application is: sample dialogue text and sample argument pairs, where each sample argument in the sample argument pair belongs to the sample dialogue
  • the text that is, the goal of model training is: how to make the initial relationship prediction model accurately predict the argument relationship between sample arguments based on the sample dialogue text.
  • the sample dialogue text may be composed of dialogue sentences generated during m rounds of dialogue.
  • each sample argument in the sample argument pair belongs to the sample dialogue text.
  • the sample argument pair includes the first sample argument and the second sample argument, that is, the first sample argument and the second sample argument. If the sample argument belongs to the sample dialogue text, the initial relationship prediction model is used to predict the argument relationship between the first sample argument and the second sample argument; optionally, the sample argument pair can also contain more than two sample argument, the initial relationship prediction model is used to predict the argument relationship between any two sample arguments in the sample argument pair.
  • the argument types of each sample argument in the sample argument pair may be speaker arguments.
  • the sample argument pair can be s 1 and N 1 , N 1 can represent any word contained in the speech content (t); or, the argument type of each sample argument in the sample argument pair is a phrase argument, that is, the sample argument pair does not contain the speaker argument, Exemplarily, the sample argument pair may be: N 1 and N 2 , N 1 and N 2 may represent any two different words contained in the speaker content (t).
  • the sample dialogue text and the sample argument pair are manually input into the computer device for subsequent training of the initial relationship prediction model; optional , the sample argument pairs can also be manually marked in the sample dialogue text, and the sample dialogue text (marked text) marked with the sample argument pairs is input into the computer equipment, and the computer equipment determines the sample dialogue text and the sample dialogue text from the marked text.
  • the computer device is provided with a sample argument pair determination method, the sample dialogue text is input into the computer device, and the computer device is based on the input sample dialogue text and the sample argument pair determination method, from the sample dialogue Sample argument pairs are extracted from the text.
  • the sample argument pair determination method can indicate the argument type of each sample argument, for example, extracting a sample argument whose argument type is speaker argument from the sample dialogue text, corresponding to any two speaker arguments Determined as a sample argument pair.
  • Step 402 perform argument relationship prediction based on the semantic information of the sample text, and determine a first loss based on the argument relationship prediction result, which is used to represent the relationship between the sample arguments.
  • the semantic information of the sample text obtained in step 401 will be shared by the three tasks.
  • the semantic information of the sample text is input into the initial relationship prediction network, and the initial relationship prediction network learns from the semantic information of the sample text The global semantics of the sample dialogue text, and the context-dependent semantics of each argument, and then based on the global semantics and context-dependent semantics, predict the relationship between the sample arguments, and further determine the first loss based on the predicted results of the argument relationship.
  • the argument relationship prediction result is the predicted argument relationship between the first sample argument and the second sample argument
  • the first The loss is determined by the predicted argument relation and the actual argument relation, which is the true argument relation between the first sample argument and the second sample argument.
  • the initial relationship prediction network is a multi-classifier
  • the prediction result of the argument relationship output by the initial relationship prediction network is: the sample argument pair corresponding argument relationship is the probability of each candidate argument relationship, and the candidate with the highest probability is selected Argument relation, which is determined as the predicted argument relation corresponding to the sample argument pair.
  • the argument relationship prediction result is the relationship between the two speaker arguments
  • Step 403 perform hidden character prediction based on the semantic information of the sample text, and determine a second loss based on the hidden character prediction result, which is used to characterize the hidden character in the sample dialogue text.
  • an initial phrase prediction network also called It is the initial speaker prediction network, which is used to predict the hidden characters in the sample dialogue text, that is, to predict the hidden speaker arguments in the sample dialogue text.
  • the computer device In order to predict the hidden characters in the sample dialogue text, there needs to be covered or hidden speaker arguments in the sample dialogue text.
  • the original dialogue text needs to be The speaker's words corresponding to the speaker's argument are masked to obtain a sample dialogue text, and then the subsequent semantic feature extraction process is performed.
  • the sample dialogue text contains hidden characters
  • the corresponding Semantic feature representation that is, the sample text semantic information contains semantic feature representation corresponding to hidden characters
  • the task principle is: if the initial phrase prediction network can distinguish which speaker What the speaker said, it means that the initial phrase prediction network can analyze the characteristics of the speaker based on the semantic information of the sample text, which further shows that the extraction of semantic information of the sample text is more accurate, which helps to predict the argument relationship related to the speaker, that is to say , the initial phrase prediction network can provide an auxiliary function for predicting the argument relationship between speaker arguments; in other words, if the initial relationship prediction model does not need to predict the argument relationship between speaker arguments, there is no need to deploy the initial phrase prediction network; if the initial relation prediction model needs to predict the argument relations among speaker arguments, then the initial phrase prediction network needs to be deployed.
  • Step 404 Predict the trigger word based on the semantic information of the sample text and the actual argument relationship, and determine the third loss based on the trigger word prediction result.
  • the actual argument relationship is the labeled argument relationship corresponding to the sample argument, and the trigger word prediction result is used for Characterize the position of trigger words in the sample dialogue text.
  • an initial trigger word prediction network is deployed in the initial relationship prediction model to predict the position of the trigger word in the sample dialogue text, that is, the initial trigger word prediction network It is necessary to identify whether each sample dialogue phrase in the sample dialogue text is a trigger word.
  • the trigger words may be dialogue phrases such as "mother” and "daughter" appearing in the sample dialogue text.
  • the computer device obtains the actual argument relationship corresponding to the sample argument pair and uses the actual Meta relations and sample text semantic information are jointly input into the initial trigger word prediction network, so that the initial trigger word prediction network can predict whether each sample dialogue phrase is a trigger word based on the sample text semantic information and the actual argument relationship, so as to obtain a representative trigger word
  • the trigger word prediction result of the position in the sample dialogue text and then determine the third loss based on the trigger word prediction result and the actual trigger word information.
  • Step 405 based on the first loss, the second loss and the third loss, train an initial relationship prediction model to obtain a dialogue relationship prediction model.
  • the proposed relationship extraction task, speaker prediction task, and trigger word prediction task are jointly trained in a multi-task joint learning manner, that is, according to the first loss (argument relationship prediction loss) , the second loss (speaker prediction loss) and the third loss (trigger word prediction loss), establish a joint loss, and then train the initial relationship prediction model based on the joint loss, after multiple rounds of repeated training until the loss converges, the dialog relationship prediction model is obtained .
  • the speaker prediction task can provide assistance for the relationship prediction task from the perspective of speaker characteristics
  • the trigger word prediction task can provide guidance for the relationship prediction task from the perspective of trigger word characteristics.
  • the Use at least one of the speaker prediction task and the trigger word prediction task to provide guidance for the relationship prediction task may include: an initial language representation network, an initial relationship prediction network, an initial phrase prediction network, and an initial trigger word The prediction network, correspondingly, during the model training process, based on the first loss (generated by the initial relationship prediction network), the second loss (generated by the initial phrase prediction network) and the third loss (generated by the initial trigger word prediction network) together Training an initial relationship prediction model; optionally, the initial relationship prediction model may also include: an initial language representation network, an initial relationship prediction network, and an initial phrase prediction network.
  • the initial relationship prediction model may also include: an initial language representation network, an initial relationship prediction network, and an initial trigger word prediction network.
  • the model training process based on the first loss and the third loss jointly train the initial relation prediction model.
  • an initial trigger word prediction network and an initial phrase prediction network will be added to assist training, but in the model application stage, the above two subtasks can be omitted, and only the relationship prediction task and semantic feature extraction task will be retained. , that is, the dialogue relationship prediction model can only include the language representation network and the relationship prediction network.
  • the initial phrase prediction network predicts the hidden speaker arguments in the sample dialogue text
  • the trigger word prediction network predicts the trigger words that can guide the argument relationship in the sample dialogue text, so as to introduce additional second loss and third loss in the loss, and provide auxiliary information for the argument relationship prediction: speaker characteristics and Trigger word features, so that the initial relationship prediction model can learn effective information that is more conducive to predicting the argument relationship, thereby improving the prediction accuracy of the argument relationship.
  • FIG. 5 is a flow chart of a method for processing a dialog relationship provided by another exemplary embodiment of the present application.
  • the dialog relationship processing method includes the following steps:
  • Step 501 Input the sample dialogue text and the sample argument pair into the initial relationship prediction model, and extract the semantic features of the sample dialogue text and the sample argument pair in the initial relationship prediction model to obtain the semantic information of the sample text.
  • the sample argument pair includes a first sample argument and a second sample argument.
  • the computer equipment obtains training samples
  • the training samples include sample dialogue texts and sample argument pairs in the sample dialogue texts, wherein the arguments refer to the part-of-speech components with thematic roles, which are matched with predicates
  • the nouns of can be called arguments
  • the sample dialogue text can be any text data including at least two arguments
  • the sample argument pair can be any two arguments in the sample dialogue text, for example, by the sample dialogue text An argument pair consisting of any two speakers in , or an argument pair consisting of any two items in the sample dialogue text, or an argument consisting of a speaker and an item in the sample dialogue text Equivalent, no limitation here.
  • the obtained sample dialogue text is "Speaker 1: 'Uh, like, could these margaritas be any stronger? Hey, Chandler.', Speaker 2: 'Hello, Mr. Bing', Speaker 3: '...', "
  • the sample argument pair can be an argument pair composed of any two speakers in the sample dialogue text, such as " ⁇ Speaker 1, Speaker 2 ⁇ , ⁇ Speaker 1, Speaker 3 ⁇ and ⁇ Speaker 2, Speaker 3 ⁇ , etc.” can also be an argument pair composed of a speaker and an item, such as " ⁇ Speaker 1, margaritas ⁇ , ⁇ Speaker 2, margaritas ⁇ , etc.”, which is not limited here.
  • the sample argument pair can have different argument types, including but not limited to same-speaker argument, same-item argument, speaker-and-item argument type, and so on.
  • the initial relationship prediction model in the process of model training, can be trained based on different argument types at the same time, and a dialogue relationship prediction model that can detect different argument types can be obtained to improve the generalization of the trained dialogue relationship prediction model Or, based on different argument types, the initial relationship prediction model can be trained separately to obtain multiple dialogue relationship prediction models.
  • the argument types detected by different dialogue relationship prediction models are different, so as to improve the dialogue relationship prediction obtained by training. model accuracy.
  • the number of detection argument types required by a model can be determined based on needs, taking comprehensive consideration of the generalization and accuracy of the model, and there is no limitation here.
  • the computer device can input the acquired sample dialogue text and sample argument pairs in the sample dialogue text into the initial relationship prediction model, and perform semantic feature extraction on the sample dialogue text and sample argument pairs in the initial relationship prediction model, and obtain The sample text semantic information corresponding to the sample dialogue text.
  • FIG. 6 is a model architecture diagram of an initial relationship prediction model provided by an exemplary embodiment of the present application.
  • the computer device obtains a training sample 601, and the training sample 601 may include sample dialogue text and sample dialogue text
  • the sample argument pair in the sample dialogue text and the sample argument pair are input into the initial relationship prediction model 602, and based on the shared encoder in the initial relationship prediction model 602, the sample dialogue text and the sample argument pair are data preprocessed and hidden.
  • the initial relationship prediction model 602 can perform relationship prediction tasks, phrase prediction tasks, and trigger word prediction tasks; optionally, in the training phase, joint training is performed based on the relationship prediction tasks, phrase prediction tasks, and trigger word prediction tasks; and in the test phase and application stage, the relation prediction results can be obtained only by using the shared encoder and relation prediction task.
  • data preprocessing needs to be performed on the sample dialogue text and sample argument pairs for the subsequent semantic feature extraction process, wherein the data preprocessing
  • the process can roughly include two steps: a sample data splicing process and a sample symbol replacement process.
  • the initial relationship prediction model may also include a data preprocessing module for performing data preprocessing on the sample dialogue text and sample argument pairs.
  • step 501 may further include steps 501A to 501C.
  • step 501A in the initial relationship prediction model, the sample dialogue text and the sample argument pairs are spliced based on the sample text splicing symbols to generate sample spliced text data.
  • sample dialogue text and sample argument pairs need to be input into the initial relationship prediction model, correspondingly, the sample dialogue text and sample argument pairs need to be spliced into a set of text data.
  • the computer device After the sample dialogue text and the sample argument pairs are acquired, the sample dialogue text and the sample argument pairs are spliced based on the sample text splicing symbols.
  • sample text concatenation may include sample global semantics and sample separators, wherein the sample global semantics is used to refer to the complete sample text data (including sample dialogue text and sample argument pairs) input to the initial relationship prediction model, and the sample The delimiter is used to separate sample dialog text and sample argument pairs.
  • sample global semantic symbol can be expressed as [CLS]
  • sample separator can be expressed as [SEP]; optionally, the sample global semantic symbol and sample separator are not limited to the above symbols, and other special symbols can also be used Symbol, which is not limited by the embodiment of the present application.
  • the sample argument pair contains the first For this argument a 1 and the second sample argument a 2 , the sample global semantic symbol is [CLS], and the sample separator is [SEP].
  • the sample generated after splicing the sample dialogue text and the sample argument pair The concatenated text data can be expressed as: [CLS]s 1 : t 1 , s 2 : t 2 , s 3 : t 3 ..., s m : t m [SEP]a 1 [SEP]a 2 [SEP].
  • the computer device can splice the sample dialogue text and sample argument pairs based on the sample text splicing symbols in the initial relationship prediction model to generate sample splicing text data, wherein the sample text splicing symbols can include sample global semantic symbols and sample Separators, etc.
  • the computer device can splice the sample dialogue text and the sample argument pair, insert the sample text splicing character into the spliced sample obtained by splicing, and obtain the sample splicing text data, wherein, the sample text splicing character and the The insertion method of the sample text concatenation can be determined based on the initial language representation network in the initial relationship prediction model, that is, different initial language representation networks have different sample splicing specifications.
  • the corresponding sample The splicing characters include [CLS] and [SEP].
  • the splicing character insertion method is as shown in the above embodiment; the sample text splicing character can be any special character or special string, which is not limited here.
  • the sample global semantic character Used to refer to the global representation of the sample dialogue text and the sample argument pair, and the sample separator is used to segment the sample dialogue text and the sample argument pair.
  • sample spliced text data is obtained, and the sample spliced text data can be "[CLS]Tx[SEP]a 1 [SEP]a 2 [SEP]", wherein it is assumed that the sample dialogue text Tx is "s 1 : t 1 s 2 : t 2 ...s u : t u ”, at this time, the sample spliced text data can be expressed as "[CLS]s 1 : t 1 s 2 : t 2 ...s u : t t t t ”,
  • Step 501B replace the first sample argument in the sample spliced text data with the first sample argument symbol, replace the second sample argument in the sample spliced text data with the second sample argument symbol, and generate sample text sequence data.
  • the training task of the initial relationship prediction model is to predict the argument relationship between the first sample argument and the second sample argument, in order to highlight the first sample argument and the second sample argument in the input sample splicing text data Sample arguments, so that the initial relationship prediction model (mainly the initial language representation network in the initial relationship prediction model) pays more attention to the sample arguments in the sample spliced text data.
  • special symbols are used to replace samples Splicing the sample arguments in the text data, this special symbol can mark the position of the sample arguments in the sample dialogue text, so that the initial language representation network can better extract the contextual semantic information related to the sample arguments, and at the same time, the The special symbols are also related to the argument type of the sample argument itself, so that the initial language representation network can extract the features related to the sample argument type, and further improve the information richness in the process of semantic feature extraction.
  • the special symbol is specially used to represent the speaker argument, in other words, when the sample argument pair contains a speaker-type sample argument, the sample argument in the sample concatenated text data is replaced;
  • the samples are concatenated with the first sample argument in the text data replacing with the first sample argument symbol, and replacing the second sample argument in the sample spliced text data with the second sample argument symbol, thereby generating sample text sequence data.
  • the sample spliced text data contains the sample dialogue text part and the sample argument pair part, when replacing, both the first sample argument and the second sample argument contained in the sample dialogue text part need to be replaced. To replace, it is also necessary to replace the first sample argument and the second sample argument of the sample argument pair.
  • the sample spliced text data is: [CLS]s 1 : t 1 , s 2 : t 2 , s 3 : t 3 ..., s m : t m [SEP]a 1 [SEP]a 2 [SEP], and both a 1 and a 2 are sample arguments of the speaker type, a 1 represents s 1 , and a 2 represents s 2 , which are used to predict the relationship between speaker 1 and speaker 2 in the sample dialogue text relationship; the first sample argument symbol is S 1 , the second sample argument symbol is S 2 , the sample spliced text data is replaced based on the first sample argument symbol and the second sample argument symbol, and the obtained sample text Sequence data can be: [CLS] S 1 : t 1 , S 2 : t 2 , s 3 : t 3 . . . , s m : t m [SEP] S 1 [SEP] S 2 [S
  • each sample argument in order to further emphasize the position of the sample argument pair in the sample spliced text data, each sample argument can also be marked with a start argument symbol and an end argument symbol position in the sample concatenated text data.
  • the start argument symbol can be represented by [B]
  • the end argument symbol can be represented by [E]; optionally, the start argument symbol and the end argument symbol can also use other forms of symbols, and this application implements Examples are not limited to this.
  • the sample spliced text data is: [CLS]s 1 : t 1 , s 2 : t 2 , s 3 : t 3 ..., s m : t m [SEP]a 1 [SEP]a 2 [SEP], after replacing the argument symbol and adding the start argument symbol and the end argument symbol, the obtained sample text sequence data can be: [CLS][B]S 1 [E]:t 1 , [B] S 2 [E]: t 2 , s 3 : t 3 ..., s m : t m [SEP] S 1 [SEP] S 2 [SEP].
  • the computer device may replace the first sample argument in the sample spliced text data with the first sample argument symbol, and replace the second sample argument in the sample spliced text data with the second sample argument symbol, Generate sample text sequence data.
  • the position of the sample argument in the sample spliced text data can be marked based on the start argument symbol and the end argument symbol.
  • the generation process of the sample text sequence data can be referred to as shown in formula 1:
  • the first sample argument can also be marked based on the start argument symbol [B] and the end argument symbol [E]
  • the position of the element in the sample spliced text data that is, replace the i-th character in the sample spliced text data with [B][S 1 ][E]; if the i-th character in the sample spliced text data is Two-sample argument (i.e.
  • sample spliced text data is generated, it is also possible to directly input the sample spliced text data into the initial language representation network for semantic feature extraction without performing sample symbol replacement on the sample spliced text data.
  • the above-mentioned process of generating sample text sequence data based on the sample dialogue text and the sample argument pairs in the sample dialogue text can be regarded as a data preprocessing process for the sample dialogue text and the sample argument pairs in the sample dialogue text.
  • Step 501C performing semantic feature extraction on sample text sequence data to obtain sample text semantic information.
  • sample text sequence data obtained after data preprocessing can be input into the initial language representation network, and the semantic features of the sample text sequence data can be extracted by the initial language representation network , and then get the relevant sample text semantic information.
  • sample text sequence data contains a variety of data, such as sample dialogue phrases (words contained in the sample dialogue text), and sample argument pairs, in the process of semantic feature extraction, different types of data will be extracted for different types of data. associated semantic representation.
  • the process of extracting semantic features from sample text sequence data may include step 1 and step 2.
  • Step 1 Perform hidden layer feature extraction on the sample global semantic symbols, N sample dialogue phrases, first sample argument symbols, and second sample argument symbols in the sample text sequence data to obtain the samples corresponding to the sample global semantic symbols
  • the sample text sequence data includes sample global semantic symbols, the first sample argument symbol, the second sample argument symbol, and N sample dialogue phrases, and the sample dialogue phrases are each sample that composes the sample dialogue text in the sample text sequence data Word or sample character.
  • the sample text sequence data is: [CLS][B]S 1 [E]: t 1 , [B]S 2 [E]: t 2 , s 3 : t 3 ..., s m : t m [SEP]S 1 [SEP]S 2 [SEP]
  • the sample dialogue phrase is [B]S 1 [E]: t 1 , [B]S 2 [E]: t 2 , s 3 : t 3 ..., s m : words or phrases contained in t m .
  • the purpose of semantic feature extraction is to extract global semantic representation and contextual semantic representation.
  • the initial language representation network recognizes the input complete sample data through sample global semantic symbols, and the sample The global semantic feature extraction is performed on the text sequence data to obtain the global hidden state of the sample (that is, the global semantic representation); and, the initial language representation network also extracts the context semantic feature of the sample argument pairs in the sample text sequence data, and obtains the same as the first The hidden state of the first initial sample argument corresponding to this argument, and the hidden state of the second initial sample argument corresponding to the second sample argument; at the same time, the initial language representation network also Dialogue phrases are subjected to contextual semantic feature extraction to obtain hidden states of N sample phrases corresponding to each sample dialogue phrase.
  • the computer device may perform feature extraction (ie hidden layer feature extraction) on the sample text sequence data to obtain sample text semantic information corresponding to the sample dialogue text.
  • the sample text concatenation symbol includes sample global semantic symbols;
  • the sample text sequence data includes sample dialogue sequence data corresponding to the sample dialogue text, and
  • the sample dialogue sequence data includes N sample dialogue phrases, where N is a positive integer.
  • the computer equipment When the computer equipment performs feature extraction on the sample text sequence data and obtains the sample text semantic information corresponding to the sample dialogue text, specifically, the sample global semantic symbols, N sample dialogue phrases, and the first sample argument in the sample text sequence data symbol and the second sample argument symbol are respectively extracted from the hidden layer features, and the sample global hidden state corresponding to the sample global semantic symbol, the sample phrase hidden state corresponding to the N sample dialogue phrases, and the first sample argument symbol corresponding to the first sample argument symbol are obtained.
  • the global hidden state of the sample, the hidden state of N sample phrases, the hidden state of the first initial sample argument and the hidden state of the second initial sample argument are determined as the sample text semantic information corresponding to the sample dialogue text.
  • Step 2 Determine the sample global hidden state, the N sample phrase hidden states, the first initial sample argument hidden state and the second initial sample argument hidden state as the sample text semantic information corresponding to the sample dialogue text.
  • the semantic information of the sample text includes the global semantic representation and the context semantic representation, wherein the global semantic representation is the global hidden state of the sample, and the context semantic representation is the hidden state of N sample phrases, the first initial sample argument hidden state and the second initial sample Argument hidden state; that is, the computer device determines the extracted sample global hidden state, N sample phrase hidden state, first initial sample argument hidden state, and second initial sample argument hidden state as sample text semantic information.
  • the computer device can obtain the sample global relationship between the N sample dialogue phrases, the first sample argument symbol and the second sample argument symbol and the sample global semantic symbol respectively based on the initial language representation network, and the sample global relationship Perform feature fusion to generate sample global hidden states corresponding to sample global semantic symbols; perform hidden layer feature extraction on N sample dialogue phrases, first sample argument symbols and second sample argument symbols respectively, and obtain N sample dialogue phrases respectively corresponding to the hidden state of the sample phrase, the hidden state of the first initial sample argument corresponding to the first sample argument symbol, and the hidden state of the second initial sample argument corresponding to the second sample argument symbol.
  • the initial relationship prediction model includes a language representation task 701 , a relationship prediction task 702 , a phrase prediction task 703 and a trigger word prediction task 704 .
  • the computer equipment obtains a training sample 7011, the training sample 7011 includes a sample dialogue text 701a and a sample argument pair 701b, if the sample dialogue text 701a is "a 1 : Mom! a 2 : Sweetie... ”, the sample argument pair 701b includes the first sample argument a 1 and the second sample argument a 2 .
  • sample data preprocessing on the training sample 7011 (that is, perform sample data splicing processing and sample symbol replacement processing) to obtain sample text sequence data 7012 (the start argument symbol and the end argument symbol are not shown in this embodiment); wherein, the data
  • the preprocessing process is: splicing the sample dialogue text 701a and the sample argument pair 701b based on the splicing characters of the sample text to generate sample spliced text data, which may be "[CLS] a 1 : Mom! a 2 : Sweetie ...[SEP] a 1 [SEP] a 2 [SEP]".
  • sample text sequence data 7012 is "[CLS]S 1 :Mom! S 2 :Sweetie...[SEP]S 1 [SEP]S 2 [SEP]".
  • the semantic feature extraction process is as follows: perform hidden layer feature extraction (semantic feature extraction) on the sample text sequence data 7012 to obtain the sample text semantic information 7013 corresponding to the sample text sequence data 7012; further, based on the sample global semantic symbol "[CLS]” Perform global semantic feature extraction on the sample text sequence data 7012 to obtain the sample global hidden state "h [CLS] " corresponding to the sample text sequence data 7012; Feature extraction, to obtain the hidden states of the sample phrases corresponding to the N sample dialogue phrases, that is, to obtain the hidden states of the N sample phrases, for example: the hidden state of the sample phrases corresponding to the sample dialogue phrase "S 1 " (Subscript 1 represents the first first sample argument that appears in the sample text sequence data 7012), the sample phrase hidden state "h i " corresponding to the sample dialogue phrase "Mom” (i represents the N sample dialogue phrase The i-th sample dialogue phrase), the hidden state of the sample phrase corresponding to the sample dialogue phrase "S 2 " At the same time, semantic feature extraction is performed on the sample argument pair
  • Step 502 predicting the predicted argument relationship among the sample arguments based on the semantic information of the sample text.
  • the context semantic information corresponding to the sample argument can be based on The semantic information of the sample text is used to predict the argument relationship between the sample arguments, that is, to obtain the predicted argument relationship between the first sample argument and the second sample argument.
  • sample text semantic information related to the sample argument which can be known from the sample text sequence data
  • the sample argument appears in the sample dialogue text and the sample argument pair.
  • semantic representation corresponding to each sample argument in the sample dialogue text and each sample argument in the sample argument pair from the semantic information of the sample text Corresponding semantic representation.
  • the process of determining the semantic information of the sample text related to the sample argument may include: the process of determining the semantic representation corresponding to the first sample argument , and determine the semantic representation process corresponding to the second sample argument.
  • step 502 may include steps 502A-502F.
  • Step 502A Obtain at least one first sample phrase hidden state corresponding to the first sample argument symbol from the N sample phrase hidden states included in the sample text semantic information.
  • the process of determining the semantic representation related to the first sample argument is as follows: Obtain N sample phrase hidden states from the sample text semantic information, and determine the first sample argument symbol (first sample argument) corresponding to at least one first sample phrase hidden state; at the same time, the first initial sample argument hidden state is obtained from the sample text semantic information, and then based on the first sample phrase hidden state and the first initial sample theory The meta-hidden state determines the semantic representation corresponding to the first sample argument.
  • step 502B the maximum pooling process is performed on each hidden state of the first sample phrase and the hidden state of the first initial sample argument to obtain the hidden state of the first sample argument corresponding to the symbol of the first sample argument.
  • the hidden states of the first sample phrases corresponding to multiple first sample arguments will also be correspondingly extracted, in order Integrating multiple first sample phrase hidden states and first initial sample argument hidden states, in a possible implementation manner, performing maximum pooling on each first sample phrase hidden state and first initial sample argument hidden state The hidden state of the first sample argument corresponding to the symbol of the first sample argument is obtained.
  • the process of determining the hidden state of the first sample argument is: obtaining the semantic feature representation corresponding to the symbol of the first sample argument (including the first sample phrase extracted from the sample dialogue text part) from the semantic information of the sample text Hidden state, and the hidden state of the first initial sample argument partially extracted by the sample argument pair), after performing maximum pooling processing on all semantic feature representations, the hidden state of the first sample argument can be obtained.
  • Step 502C obtaining at least one second sample phrase hidden state corresponding to the second sample argument symbol from the N sample phrase hidden states.
  • the process of determining the semantic representation related to the second sample argument is: obtaining N sample phrase hidden states from the sample text semantic information, and determining the second sample argument symbol (second sample theory) from the N sample phrase hidden states element) corresponding to at least one second sample phrase hidden state; at the same time, obtain the second initial sample argument hidden state from the sample text semantic information, and then based on the second sample phrase hidden state and the second initial sample argument hidden state, determine The semantic representation corresponding to the second sample argument.
  • Step 502D performing maximum pooling processing on each second sample phrase hidden state and the second initial sample argument hidden state to obtain the second sample argument hidden state corresponding to the second sample argument symbol.
  • the hidden state of the second sample phrase corresponding to multiple second sample arguments will also be correspondingly extracted.
  • the process of determining the hidden state of the second sample argument is: obtaining the semantic feature representation corresponding to the symbol of the second sample argument from the semantic information of the sample text (including the hidden state of the second sample phrase extracted from the sample dialogue text, And sample arguments partially extracted second initial sample argument hidden state), after performing maximum pooling processing on all semantic feature representations, the second sample argument hidden state can be obtained.
  • Step 502E concatenate the sample global hidden state, the first sample argument hidden state and the second sample argument hidden state to obtain sample hidden state information.
  • the sample global hidden state is obtained from the sample text semantic information, and then based on the sample global hidden state, the first sample argument hidden state and the second sample argument hidden state to jointly perform argument relationship prediction.
  • the initial relationship prediction network in order to input the three hidden state information into the initial relationship prediction network together, it is necessary to concatenate the sample global hidden state, the first sample argument hidden state, and the second sample argument hidden state to obtain the sample hidden state information , and then output the sample hidden state information to the initial relationship prediction network to predict the argument relationship.
  • sample global hidden state can be expressed as h [CLS]
  • first sample argument hidden state can be expressed as The hidden state of the second sample argument can be expressed as Then the sample hidden state information can be expressed as
  • Step 502F predicting the predicted argument relationship between the first sample argument and the second sample argument based on the hidden state information of the sample.
  • the sample hidden state information can be input into the initial relationship prediction network, and the initial relationship prediction network can predict the first sample argument and The predicted argument relationship between the arguments of the second sample.
  • the relationship prediction task 702 based on the sample hidden state information 7021 (including the sample global hidden state h [CLS] , the first sample argument hidden state and the second sample argument hidden state ) to predict the argument relationship, and obtain the predicted argument relationship corresponding to the first sample argument and the second sample argument.
  • the sample hidden state information 7021 including the sample global hidden state h [CLS] , the first sample argument hidden state and the second sample argument hidden state
  • the initial relationship prediction model includes an initial relationship prediction network
  • the N sample dialogue phrases include the first sample argument symbol, the second sample argument symbol, and the first sample argument and the second Additional sample dialog phrases other than sample arguments.
  • the computer device can predict the network based on the initial relationship, from N sample phrases hidden in the semantic information of the sample text In the state, obtain the first sample phrase hidden state corresponding to the first sample argument symbol, perform maximum pooling processing on the first sample phrase hidden state and the first initial sample argument hidden state, and obtain the first sample theory
  • the hidden state of the first sample argument of the meta symbol; the hidden state of the second sample phrase corresponding to the second sample argument symbol is obtained from the hidden states of N sample phrases, and the hidden state of the second sample phrase and the second initial sample argument
  • the hidden state is subjected to maximum pooling processing to obtain the second sample argument hidden state of the second sample argument symbol.
  • the hidden state of the first sample argument and the hidden state of the second sample argument can also be generated in the shared encoder (or the initial language representation network in the above embodiment), and the initial relationship prediction model
  • the hidden state of the first sample argument and the hidden state of the second sample argument obtained by the shared encoder are input into the initial relationship prediction network.
  • the computer device can splicing the sample global hidden state, the first sample argument hidden state and the second sample argument hidden state to obtain the sample hidden state information; based on the sample hidden state information, predict the first sample argument and The predicted argument relationship between the arguments of the second sample.
  • the computer device can also perform semantic enhancement on the hidden state information of the sample, and then carry out the argument relationship based on the enhanced semantic information of the sample predict.
  • step 502F may also include the following steps 3 to 5.
  • Step 3 Semantic enhancement is performed on the hidden state information of the sample to obtain the enhanced semantic information of the sample.
  • a network with semantic enhancement function can be used to enhance the semantics of the hidden state information of the sample, such as the highway neural network, or other neural networks with semantic enhancement function; in other words, the initial relationship prediction model can also include highway A neural network, the expressway neural network is used to semantically enhance the extracted semantic information.
  • the sample hidden state information is input into the expressway neural network, and the expressway neural network performs semantic enhancement on it to extract deeper semantic features, obtain sample enhanced semantic information, and then enhance Semantic information for argument relationship prediction.
  • Step 4 Determine the sample relationship prediction probabilities of M candidate argument relationships corresponding to the first sample argument and the second sample argument based on the sample enhanced semantic information, where M is a positive integer.
  • M kinds of candidate argument relationships are preset, and the task of the initial relationship prediction network is to enhance semantic information based on the input samples, predict the first sample argument and the second sample argument Predicted argument relation of element, the predicted probability of sample relation belonging to M kinds of candidate argument relations.
  • the sample enhanced semantic information is input into the initial relationship prediction network, and the initial relationship prediction network predicts the first Predicted probabilities of sample relations for which one sample argument and a second sample argument belong to various candidate argument relations.
  • Step 5 Determine the candidate argument relationship corresponding to the maximum sample relationship prediction probability as the predicted argument relationship between the first sample argument and the second sample argument.
  • the candidate argument relationship corresponding to the maximum sample relationship predicted probability is determined as the predicted argument between the first sample argument and the second sample argument by default relation.
  • semantic enhancement is performed on the sample hidden state information to obtain sample enhanced semantic information.
  • the computer device can perform semantic enhancement on the hidden state information of the sample based on the fusion network, and the fusion network can be any network that can perform semantic enhancement, such as a network including one or at least two highway networks (highway network), etc. , based on the sample enhanced semantic information, predict the sample relationship prediction probabilities of the M candidate argument relationships corresponding to the first sample argument and the second sample argument, and determine the candidate argument relationship with the largest sample relationship prediction probability as The predicted argument relationship between the first sample argument and the second sample argument, M is a positive integer.
  • the sample hidden state information is directly input into the initial relationship prediction network, and the initial relationship prediction network is based on the global information contained in the sample hidden state information. Semantic information, contextual semantic information, and argument-related semantic information, etc., predict the predicted argument relationship between the first sample argument and the second sample argument.
  • Step 503 generating a first loss according to the actual argument relationship and the predicted argument relationship between the sample arguments.
  • the computer device may generate the first loss according to the actual argument relationship and the predicted argument relationship between the first sample argument and the second sample argument, and the first loss may be the actual argument relationship and the predicted argument relationship Binary cross-entropy loss, logarithmic loss, or square loss between relationships is not limited here.
  • the first sample phrase hidden state corresponding to the first sample argument symbol S 1 is obtained, namely etc., for the hidden state of the first sample phrase and the hidden state of the first initial sample argument Perform maximum pooling to obtain the first sample argument hidden state of the first sample argument symbol
  • the second sample phrase hidden state corresponding to the second sample argument symbol S 2 from N sample phrase hidden states, namely etc., for the second sample phrase hidden state and the second initial sample argument hidden state Perform maximum pooling processing to obtain the second sample argument hidden state of the second sample argument symbol
  • relation prediction task 702 for the sample global hidden state h [CLS] , the first sample argument hidden state and the hidden state of the second sample argument Splicing is performed to obtain the sample hidden state information 7021, and based on the sample hidden state information 7021, the predicted argument relationship between the first sample argument and the second sample argument is predicted.
  • the sample hidden state information 7021 can be recorded as h
  • the sample global hidden state represents the global semantic information in the sample text sequence data, and integrates the relationship between each phrase in the sample text sequence data and the sample global semantic symbols.
  • the hidden state representation of the first sample argument and the second The local semantic information related to a sample argument, the hidden state of the second sample argument represents the local semantic information related to the second sample argument, and the global hidden state h [CLS] of the sample, the hidden state of the first sample argument and the hidden state of the second sample argument Splicing is performed to obtain the sample hidden state information 7021, so that the sample hidden state information can include global semantic information and local semantic information in the sample text sequence data, and the features in the sample text sequence data can be extracted more comprehensively, thereby improving the relationship forecast accuracy.
  • Step 504 obtaining hidden characters in the sample dialogue text.
  • the sample dialogue text input to the initial relationship prediction model is the dialogue text after the cover or hidden processing, that is to say, there is the original dialogue text, for After the original dialogue text is hidden, a sample dialogue text can be obtained.
  • the process of concealing the original dialogue text to obtain the sample dialogue text may include steps six to eight.
  • Step 6 Obtain the original dialogue text and sample argument pairs.
  • the original dialogue text the dialogue text that has not been hidden.
  • Each sample argument in a sample-argument pair also belongs to the original dialogue text.
  • Step 7 In response to the argument type of at least one sample argument in the sample argument pair being a speaker argument, determine hidden characters based on the sample argument pair.
  • the speaker prediction task (initial phrase prediction network) is used to predict the masked speaker, so that the initial relationship prediction model can learn the characteristics of the speaker, and then assist the speaker-related parameter relationship prediction, the corresponding need is based on the sample theory
  • the speaker argument in the sample argument pair is determined to cover up the corresponding characters in the original dialogue text. Therefore, in a possible implementation, when there is at least one speaker argument in the sample argument pair, it can be based on the speaker argument Meta determines the hidden characters that need to be hidden in the original sample dialogue text.
  • any sample argument can be randomly selected, and the corresponding phrase of the sample argument in the original dialogue text Determined as a hidden character; optionally, if the sample argument pair only contains a single speaker argument, directly determine the phrase corresponding to the speaker argument in the original dialogue text as a hidden character.
  • Step 8 Perform hidden processing on the original dialogue text based on the hidden characters to obtain a sample dialogue text.
  • the phrase corresponding to the hidden character in the original dialogue text may be hidden, and then the hidden original dialogue text may be determined as a sample dialogue text, Perform subsequent semantic feature extraction, argument relationship prediction, trigger word prediction, hidden character prediction and other tasks.
  • the sample argument pair is (s 1 , s 2 )
  • each sample argument contained in the sample argument pair is a speaker argument
  • s 1 can be determined as a hidden character
  • the hiding process may be to replace the hidden character with other meaningless characters, or to garble the hidden character, or to use other hiding methods, which is not limited in this embodiment of the present application.
  • sample arguments in the original dialogue text also need to participate in the subsequent argument relationship prediction process, when the original dialogue text is hidden based on hidden characters, all hidden characters contained in the original dialogue text ( sample argument) for hidden processing, but randomly mask the hidden characters in the original dialogue text with a preset probability.
  • the preset probability can be 10%, that is, randomly cover 10% of the hidden characters in the original dialogue text, for example, if the original dialogue text contains 10 hidden characters, then randomly select 1 from the 10 hidden characters Hidden characters are hidden, and the sample dialogue text is obtained.
  • the initial phrase prediction network needs to predict the hidden characters that are covered or hidden in the sample dialogue text based on the semantic information of the sample text. characters, so that the hidden character prediction loss can be calculated based on the hidden character prediction results and hidden characters.
  • Step 505 Predict the predicted character corresponding to the hidden character based on the semantic information of the sample text.
  • the semantic information of the sample text output by the initial language representation network is shared by the initial relationship prediction network, the initial phrase prediction network, and the initial trigger word prediction network respectively. Therefore, in a possible implementation, the initial phrase The prediction network can also predict the hidden characters hidden in the sample dialogue text based on the semantic information of the sample text, and obtain the predicted characters.
  • step 505 may include step 505A and step 505B.
  • Step 505A determine the mask hidden state corresponding to the hidden character from the semantic information of the sample text, and the mask hidden state is used to represent the semantic information corresponding to the hidden character in the sample dialogue text.
  • Step 505B based on the hidden state of the mask, predict the predicted character corresponding to the hidden character.
  • the predicted character can be predicted based on the mask hidden state.
  • the semantic information of the sample text output by the initial language representation network can be obtained.
  • the semantic information of the sample text contains "C"
  • the hidden state of the mask corresponding to this hidden character Then the mask hides the state Input the initial phrase prediction network, perform hidden character prediction, and obtain predicted characters.
  • the phrase prediction task 703 the mask hidden state corresponding to the hidden character is obtained from the sample text semantic information 7013 hide state Input the initial phrase prediction network to get the hidden state of the mask The corresponding predicted character.
  • Step 506 generating a second loss according to the hidden character and the predicted character.
  • a hidden character prediction loss i.e. the second loss, is determined for updating the initial relation prediction model.
  • the computer device can generate a second loss according to the hidden character and the predicted character, and the second loss can be a binary cross-entropy loss, a logarithmic loss or a square loss between the predicted character and the hidden character, etc., and there is no limitation here .
  • Step 507 generate trigger word detection text data according to the actual argument relationship and the semantic information of the sample text.
  • the initial relationship prediction model also includes an initial trigger word prediction network, wherein the trigger word refers to a word in the text that can clearly indicate the relationship between arguments, and the trigger word prediction network is used to predict based on the semantic information of the sample text. Trigger words in sample dialogue text.
  • trigger word detection text data can be generated based on actual argument relationships and sample text semantic information, and the trigger word detection text data can be input into the initial trigger word prediction network , the trigger words in the sample dialogue text are predicted by the initial trigger word prediction network.
  • step 507 may include step 507A and step 507B.
  • Step 507A determine the argument relation vector corresponding to the actual argument relation.
  • the actual argument relationship is also converted into the form of feature vectors, that is, The argument relation vector corresponding to the actual argument relation.
  • Step 507B splicing the argument relationship vector and the semantic information of the sample text to generate trigger word detection text data.
  • the argument relationship vector and the semantic information of the sample text can be spliced to generate trigger word detection text data, which is used in the subsequent trigger word prediction process.
  • step 507B may also include the following steps 9 to 11.
  • Step 9 Determine at least one hidden state of the sample phrase from the semantic information of the sample text, and the hidden state of the sample phrase is used to represent the semantic information corresponding to the sample dialogue phrase in the sample dialogue text.
  • sample dialogue phrase is each character contained in the sample dialogue text.
  • sample dialogue text is: a 1 :Mom! a 2 : Sweetie
  • sample dialogue phrases include: a 1 , Mom, a 2 , Sweetie, etc.
  • the semantic information corresponding to each sample dialogue phrase is determined from the sample text semantic information, and also It is the hidden state of the sample phrase, and then the hidden state of the sample phrase is spliced with the argument relationship vector, which is used to subsequently predict whether the sample dialogue phrase is a trigger word.
  • Step 10 Concatenate the argument relationship vector with the hidden state of the sample phrase to obtain the trigger word detection text corresponding to the sample dialogue phrase.
  • the hidden state of the sample phrase is spliced with the argument relationship vector to obtain the trigger word detection text corresponding to the sample dialogue phrase, that is to say, the trigger word detection text contains the argument relationship vector and a sample Phrase hidden state.
  • the sample phrase hidden state "hi” corresponding to the sample dialogue phrase "Mom” is obtained from the sample text semantic information 7013 , and the sample phrase hidden state " hi " is related to the argument relationship vector "e r " to get the trigger word detection text "[e r , h i ]" corresponding to the sample dialogue phrase "Mom”.
  • Step 11 Determine the trigger word detection text corresponding to each sample dialogue phrase as trigger word detection text data.
  • the hidden state of the sample phrase corresponding to each sample dialogue phrase is spliced with the argument relationship vector to obtain the trigger word detection text corresponding to each sample dialogue phrase , and then determine the set of trigger word detection texts as the trigger word detection text data, and input the initial trigger word prediction network together, and the initial trigger word prediction network simultaneously predicts whether each sample dialogue phrase is a trigger word.
  • the sample dialogue text contains 5 sample dialogue phrases
  • each sample The hidden state of the phrase is spliced with the argument relationship vector to obtain 5 trigger word detection texts: [er r , h 1 ], [er r , h 2 ], [er r , h 3 ], [er r , h 4 ] , [e r , h 5 ]
  • five trigger word detection texts constitute the trigger word detection text data, which are used to input the initial trigger word detection network for trigger word prediction.
  • Step 508 predicting trigger words and detecting text data corresponding to predicted sequence annotations.
  • the computer device inputs the trigger word detection text data into the initial trigger word prediction network, and the initial trigger word prediction network judges the sample dialogue based on the semantic information of each sample dialogue phrase and the actual argument relationship Whether there is a specific relationship between the phrase and the actual argument relationship, so as to determine whether the sample dialogue phrase is a trigger word.
  • step 508 may further include step 508A and step 508B.
  • step 508A the trigger word prediction is performed based on the trigger word detection text data, and the predicted phrase annotation corresponding to each sample dialogue phrase is obtained, and the predicted phrase annotation is used to represent the trigger word type to which the sample dialogue phrase belongs.
  • the BIO tagging mode is used to mark the sample dialogue text, that is to say, the output of the trigger word prediction network is the prediction composed of BIO Sequence annotation, wherein, I indicates that the sample dialogue phrase at this position is the middle character of the trigger word, B indicates that the sample dialogue phrase at this position is the initial character of the trigger word, and O indicates that the sample dialogue phrase at this position is not a trigger word .
  • BIOES labeling mode wherein, B indicates that the sample dialogue phrase at this position is the initial character of the trigger word, and I indicates that the sample dialogue phrase at this position is the middle of the trigger word character, O represents that the sample dialogue phrase at this position is not a trigger word, E represents that the sample dialogue phrase at this position is the end character of the trigger word, and S represents that this sample dialogue phrase is a single character (independent of the trigger word).
  • the trigger word detection text data is input into the initial trigger word prediction network, and the trigger word prediction is performed to obtain the predicted phrase annotations corresponding to each sample dialogue phrase, that is, to determine whether the sample dialogue phrase is a trigger word , and whether it is the start, middle, or end character of the trigger word.
  • the initial phrase prediction network will predict the probability that the sample dialogue phrase corresponds to multiple candidate phrases.
  • the candidate phrases are marked as three types: B , I, O; if the labeling mode adopts the BIOES labeling method, the candidate phrases are marked as five types: B, I, O, E, S; and then the candidate phrases corresponding to the maximum probability are marked as the predicted phrases corresponding to the sample dialogue phrases label.
  • Step 508B labeling the predicted phrase corresponding to each sample dialogue phrase, and determining it as a predicted sequence label.
  • the predicted phrase annotations corresponding to each sample dialogue phrase can constitute the predicted sequence annotation corresponding to the sample dialogue text, that is, the predicted sequence annotation is a set of each predicted phrase annotation.
  • the trigger word detection text data is: [er r , h 1 ], [er r , h 2 ]...[er r , h i ], [er r , h i+1 ], [er r , h i+2 ], input the trigger word detection text data into the initial trigger word prediction network, the obtained prediction tag sequence is: O, O...B, I, O, it can be seen that the first sample dialogue phrase, the second sample The prediction labels corresponding to the dialogue phrase and the i+2th sample dialogue phrase are all O, indicating that the first dialogue sample phrase, the second dialogue sample phrase and the i+2th sample dialogue phrase are not trigger words, and the i The prediction label corresponding to the sample dialogue phrase is B, which means that the i-th sample dialogue phrase is the initial character of the trigger word, and the corresponding prediction label of the i+1 sample dialogue phrase is I, which means that the i+1 sample dialogue phrase is the trigger word.
  • the predicted trigger word in the sample dialogue text consists of the i-th sample dialogue phrase and the i+1-th sample dialogue phrase.
  • Step 509 generate a third loss according to the actual sequence annotation and predicted sequence annotation corresponding to the trigger word detection text data.
  • the trigger word prediction loss is determined, namely The third loss is used to update the initial relation prediction model.
  • the actual sequence annotation is pre-labeled manually, and the actual sequence annotation is associated with the corresponding sample dialogue text and stored in the computer device, so that the computer device can obtain the corresponding actual sequence annotation when calculating the third loss.
  • the computer device can detect the actual sequence annotation and the predicted sequence annotation in the text data according to the trigger words to generate a third loss
  • the third loss can be binary cross entropy and logarithmic loss between the actual sequence annotation and the predicted sequence annotation Or square loss, etc., there is no limit here.
  • the trigger word detection text data is generated using the sample argument pair corresponding to the actual argument relationship; and in the model testing phase, in order to verify the initial relationship prediction model in three tasks Prediction accuracy in , use the predicted argument relationship output by the initial relationship prediction network to generate trigger word detection text data, that is, the computer device can obtain the predicted argument relationship obtained in the initial relationship prediction network, and predict the argument relationship Perform vector conversion to obtain the predicted relationship vector corresponding to the predicted argument relationship, splice the predicted relationship vector with the hidden state of N sample phrases, generate trigger word test data, predict the test sequence label in the trigger word test data, and label based on the test sequence
  • the fourth loss generated with actual sequence annotations is used to optimize the adjustment of the initial relation prediction network.
  • Step 510 Adjust the model parameters of the initial relationship prediction model according to the first loss, the second loss and the third loss to generate a dialog relationship prediction model.
  • the joint loss is determined according to the first loss, the second loss, and the third loss, and then the initial relationship prediction model is trained based on the joint loss, and the model parameters of each sub-network in the initial relationship prediction model are updated. After multiple rounds of training, until the loss converges, a dialog relationship prediction model is obtained.
  • a test sample set can be used to perform model testing on the dialogue relationship prediction model, so as to further optimize the dialogue relationship prediction model.
  • the initial phrase prediction network and the initial trigger word prediction network need to be assisted in training, and when the initial dialogue relationship prediction model is trained, in order to simplify the model, the initial The initial phrase prediction network and the initial trigger word prediction network in the dialogue relationship prediction model obtain the dialogue relationship prediction model, that is, the dialogue relationship prediction model can only include the language representation network and the relationship prediction network; optionally, the dialogue relationship prediction model Semantic augmentation networks can also be included.
  • the dialogue relationship prediction model is used to predict the target argument relationship between the target argument pairs in the target dialogue text.
  • the computer device can generate a model loss according to the first loss, the second loss, and the third loss, and adjust the model parameters of the initial relationship prediction model based on the model loss, specifically, adjust the model parameters of each network in the initial relationship prediction model , to generate a dialog relationship prediction model.
  • the first loss is denoted as L 1
  • the second loss is denoted as L 2
  • the third loss is denoted as L 3
  • the model loss is obtained based on the first loss, the second loss and the third loss
  • the model parameters of the initial relationship prediction model are adjusted based on the model loss.
  • the relationship prediction task, phrase prediction task and trigger word prediction task are jointly trained in a multi-task learning manner, so that the information of the phrase prediction task and trigger word prediction task can be provided to the relationship prediction task, that is, the The relationship prediction task can obtain the feature information of the argument based on the phrase prediction task, and the trigger word information can be obtained based on the trigger word prediction task, so that the relationship prediction task can more comprehensively obtain the information in the sample dialogue text and improve the relationship forecast accuracy.
  • multi-task learning Multi-task Learning, MTL
  • MTL Multi-task Learning
  • the sample dialogue text and the sample argument pairs in the sample dialogue text are input into the initial relationship prediction model, and feature extraction is performed on the sample dialogue text and the sample argument pairs in the initial relationship prediction model to obtain the sample dialogue text
  • the corresponding sample text semantic information includes the first sample argument and the second sample argument; predict the relationship between the first sample argument and the second sample argument based on the sample text semantic information , generate the first loss function according to the actual argument relationship and predicted argument relationship between the first sample argument and the second sample argument; obtain the hidden characters in the sample dialogue text and the sample argument pair, and predict the hidden character correspondence
  • the predicted characters generate the second loss function according to the hidden characters and predicted characters; generate the trigger word detection text data according to the actual argument relationship and the semantic information of the sample text, predict the trigger words to detect the prediction sequence annotation in the text data, and detect the text according to the trigger words
  • the actual sequence labeling and predicted sequence labeling in the data generate the third loss function; adjust the model parameters of the initial relationship prediction model according to the first loss function, the second loss function and
  • this application divides conversational relationship extraction into three related sub-tasks, which are relationship prediction task, phrase prediction task and trigger word prediction task. By combining these three sub-tasks and jointly training the model, it can fully The effective information learned from the phrase prediction task and trigger word prediction task is used, and the effective information is used to influence the relationship prediction task, so as to improve the accuracy of dialogue relationship processing.
  • the above embodiment mainly describes the training process of the initial relationship prediction model. After the training of the initial relationship prediction model is completed and the dialogue relationship prediction model is generated, the dialogue relationship prediction model can be used in different dialogue relationship prediction scenarios.
  • FIG. 8 is a flowchart of a method for processing a dialog relationship provided by an exemplary embodiment of the present application.
  • the method is applied to the computer equipment shown in FIG. 1 as an example. The method includes the following steps:
  • Step 801 input the target dialogue text and the target argument pair into the dialogue relationship prediction model, perform semantic feature extraction on the target dialogue text and the target argument pair in the dialogue relationship prediction model, and obtain the target text semantic information corresponding to the target dialogue text, the target Each target argument in an argument pair belongs to the target dialogue text.
  • the dialogue relationship prediction model can only include the language representation network and the relationship prediction network; where the language representation network is used to extract the target text semantic information corresponding to the target dialogue text, and the relationship prediction network is used to predict each target in the target argument pair
  • the target argument relationship between arguments optionally, the dialog relationship prediction model may also include a semantic enhancement network, which is used to perform semantic enhancement on the semantic information of the target text.
  • the target dialogue text and the target argument are input into the language representation network in the dialogue relationship prediction model, and the language representation network performs semantic feature extraction on it to obtain the target text semantics corresponding to the target dialogue text information.
  • the computer device splices the target dialog text and the target argument pair based on the target text splicing symbol to generate target spliced text data; replace the first target argument in the target spliced text data with the first target argument Meta symbol, replace the second target argument in the target spliced text data with the second target argument symbol to generate the target text sequence data; perform semantic feature extraction on the target text sequence data to obtain the target text semantic information corresponding to the target dialogue text .
  • the target text splicing character is the same as the sample text splicing character, but the name is different at different stages.
  • the sample text splicing character includes the sample global semantic character and the sample separator, etc., and the sample global semantic character is assumed to be [CLS] , the sample delimiter is [SEP], then the target text splicing character includes the target global semantic character and target delimiter, etc., the target global semantic character is [CLS], and the target delimiter is [SEP].
  • the target text sequence data includes target dialogue sequence data corresponding to the target dialogue text
  • the target dialogue sequence data includes v target dialogue phrases, where v is a positive integer
  • the computer device can analyze the target global semantic symbols, The v target dialog phrases, the first target argument symbol and the second target argument symbol respectively perform hidden layer feature extraction to obtain the target global hidden state corresponding to the target global semantic symbol and the target phrase hidden state corresponding to N target dialog phrases , the first initial target argument hidden state corresponding to the first target argument symbol and the second initial target argument hidden state corresponding to the second target argument symbol; the target global hidden state, v target phrase hidden states, the first The initial target argument hidden state and the second initial target argument hidden state are determined as target text semantic information corresponding to the target dialogue text.
  • the target dialogue text and the target argument pair can be input by the user, that is, the user inputs the target dialogue text and at least one target argument pair that needs to extract the argument relationship into the computer device, and the computer device obtains to the target dialogue text and the target argument pair; in other possible implementation manners, the target argument pair may need to be obtained by the computer device from the dialogue consultation information, and the dialogue consultation information may be: user consultation sentences, reading comprehension questions, etc. .
  • the computer device may directly acquire the target dialogue text provided by the target user and the target argument pairs in the target dialogue text.
  • the computer device may obtain the dialogue consultation information associated with the target dialogue text, analyze the dialogue consultation information, and extract the target argument pairs in the dialogue consultation information; for example, the acquired dialogue consultation information is "What is the relationship between Speaker 2 and Speaker 4?", analyze the dialogue consultation information, and extract the target argument pair in the dialogue consultation information, it can be obtained that the target argument pair includes the first target argument "Speaker 2" and the second target argument "Speaker 4".
  • the computer device obtains the first target argument in the dialogue consultation information, then parse the dialogue consultation information to obtain the associated argument type, obtain the argument corresponding to the association argument type from the target dialogue text, and set the The argument corresponding to the associated argument type is determined as the second target argument, wherein the associated argument type includes but not limited to character type, item type or animal type, etc., and the number of the second target argument is one or at least two indivual. Based on the dialog relationship prediction model, the target argument relationship between the first target argument and each second target argument is obtained.
  • Step 802 Predict the semantic information of the target text based on the relationship prediction network in the dialogue relationship prediction model, and obtain the target argument relationship between the first target argument and the second target argument.
  • the computer device after the computer device obtains the target text semantic information corresponding to the target dialogue text, it can determine the target global semantic information and the first target theory corresponding to the first target argument from the target text semantic information. meta-hidden state, and the second target argument hidden state corresponding to the second target argument, and determine the target global semantic information, the first target argument hidden state, and the second target argument hidden state as the target hidden state information , input the relation prediction network, predict the argument relationship by the relation prediction network, and obtain the target argument relationship between the first target argument and the second target argument.
  • the determination process of the hidden state of the first target argument can refer to the determination process of the hidden state of the first sample argument in the above embodiment
  • the determination process of the hidden state of the second target argument can also refer to the above embodiment
  • the process of determining the hidden state of the second sample argument in the second sample is not described in this embodiment of the present application.
  • the dialog relationship prediction model is obtained by training the initial relationship prediction model based on the first loss, the second loss and the third loss;
  • the initial relationship prediction model includes the initial relationship prediction network, the initial phrase prediction network and the initial trigger word prediction network;
  • the first loss function is generated by the actual argument relationship and predicted argument relationship between the first sample argument and the second sample argument, the actual argument relationship is in the initial relationship prediction network, for The first sample argument and the sample text semantic information corresponding to the sample dialogue text where the second sample argument is located;
  • the second loss is generated by hidden characters and predicted characters, and the predicted characters are in the initial phrase prediction network, It is obtained by predicting the hidden characters in the sample dialogue text;
  • the third loss is generated by the actual sequence annotation and the predicted sequence annotation in the trigger word detection text data.
  • the trigger word detection text data is in the initial trigger word prediction network, according to The actual argument relationship and the semantic information of the sample text are generated, and the predicted sequence annotation is obtained by predicting the trigger word detection text data.
  • the computer device can adjust the model parameters of the initial relationship prediction model, delete the phrase prediction task area and the trigger word prediction task area, and obtain the dialogue relationship prediction model, which can simplify the model and reduce the resources occupied by the dialogue relationship prediction model.
  • the phrase prediction task area and the trigger word prediction task area are reserved to obtain the dialogue relationship prediction model, which is convenient for further optimization of the dialogue relationship prediction model.
  • the dialogue The model structure of the relationship prediction model can be shown in FIG. 7 .
  • the computer equipment performs joint training based on the relationship prediction task, phrase prediction task and trigger word prediction task to generate a dialogue relationship prediction model; in the test phase, the relationship prediction result is obtained directly based on the relationship prediction task, The relationship prediction result is used to represent the argument relationship between target argument pairs input into the dialog relationship prediction model.
  • the computer device may generate target question-and-answer data according to the target dialogue text, dialogue consultation information, and target argument pairs, and add the target question-and-answer data to the question-and-answer database.
  • the computer device parses the dialogue consultation information to obtain the associated argument type. Specifically, if the computer device obtains the target argument relationship in the dialogue consultation information, then obtain the relevant argument type corresponding to the target argument relationship, and obtain the candidate argument corresponding to the relevant argument type from the target dialogue text, based on the dialogue
  • the relationship prediction model predicts the candidate argument relationship between the first target argument and the candidate argument, and determines the candidate argument whose candidate argument relationship is the target argument relationship as the second target argument, and the first target argument and The second target argument forms a target argument pair, wherein the second target argument is the reply data of the dialog consultation information, and the number of the candidate arguments is one or at least two.
  • the dialogue consultation information is "Who is the boss of Speaker 2? (Who is the boss of user 2)", the first target argument "Speaker 2" is obtained in the dialogue consultation information, and the obtained target argument relationship is "boss", the target argument relationship is that the associated argument type corresponding to "boss (boss)" is the character type, it is assumed that the computer device obtains the candidate arguments corresponding to the character type from the target dialogue text, and it is assumed that the candidate arguments include " Speaker 1, Speaker 3, and Speaker 4", based on the dialogue relationship prediction model, predict the relationship between the first target argument and each candidate argument, assuming that the first target argument and the candidate argument "Speaker 1"
  • the candidate argument relationship between is "subordinate (subordinate)"
  • the candidate argument relationship between the first target argument and candidate argument "Speaker 3" is "friend (friend)”
  • the candidate argument and candidate If the candidate argument relationship between the arguments "Speaker 4" is "boss", then the candidate argument "Speaker 4" is determined as the second target argument
  • this application can be applied to the information extraction business in the field of cultural tourism to optimize the relationship prediction ability, quickly extract effective knowledge information from the introduction document, and customize the privatized knowledge map; or, it can be applied to question answering system, to help improve the knowledge base of the question answering system, to ensure that the question answering system can answer user questions more effectively, and there is no limitation here.
  • the input parameters of the initial relationship prediction model may also include task type, task version, usage area, input data path, output data path, and task identifier, etc., which are not limited here.
  • task type e.g., task version, usage area, input data path, output data path, and task identifier, etc.
  • the task type (Action) is used to indicate whether the task type of this model prediction task is a training type or a test type, etc., and the task type at this time can be a training type
  • the task version (Version) is used to indicate all The version number or version generation time of the model used, etc.
  • the use region (Region) is used to indicate the list of regions, product types, or user lists that can apply the model
  • the input data path (InputDatasetPath) is used to indicate the training model, etc.
  • each input parameter in Table 1 may not be input into the initial relationship prediction model, but only used as a log record of the model training phase.
  • the output parameters of the initial relationship prediction model may also include training completion time and task request identification, etc., which are not limited here, as shown in Table 2 for details:
  • the training completion time (TimeOfTrain) is used to represent the duration used for this model prediction task completion or the time when it is completed;
  • the task request identifier (RequestId) is used to represent the request identifier requesting this model prediction task ( ID) is the unique identifier of the request, which can be used to indicate the scene change of the task, the user ID of the request or the number of requests, etc.
  • the input parameters of the dialog relationship prediction model may also include task type, task version, use area, input data path, output data path, and task identification, etc., which are not limited here. See Table 3:
  • the task type (Action) is used to indicate whether the task type of this model prediction task is a training type or a test type, etc., and the task type at this time can be a test type or a prediction type; Version is used to indicate the used The version number or version generation time of the model; Region is used to indicate the list of regions, product types, or users that can apply the model; the input data path (InputDatasetPath) is used to indicate the data used to test the model or use the model, etc.
  • the storage location path of the data such as the storage path of the target dialog text
  • the output data path (OutputModelPath) is used to indicate the storage location path of the trained model, etc.
  • the task identifier (ProjectId) is used to indicate the identifier of this model prediction task (ID).
  • each input parameter in Table 3 may not be input into the dialog relationship prediction model, but only used as a log record in the model training phase.
  • the output parameters of the dialog relationship prediction model may also include output result data paths and task request identifiers, etc., which are not limited here, as shown in Table 4 for details:
  • the output result data path (OutputDatasetPath) is used to indicate the storage location path of the data obtained in this model prediction task, such as the target argument relationship, etc., and the path can be a location path in the blockchain network , or the storage path in the computer device, or the storage path indicated by the cloud storage technology, etc., there is no limitation here;
  • the task request identifier (RequestId) is used to indicate the request identifier (ID) for requesting this model prediction task , is the unique identifier of the request, which can be used to indicate the scene change of the task, the ID of the user who initiated the request, or the number of requests, etc.
  • FIG. 9 is a schematic structural diagram of an apparatus for processing dialogue relations provided by an exemplary embodiment of the present application.
  • the dialog relationship processing device can be a computer program (including program code, etc.) running in a computer device, for example, the dialog relationship processing device can be an application software; the device can be used to execute the method in the embodiment of the present application corresponding steps.
  • the dialog relationship processing device 900 may include: a model input module 11, a semantic information extraction module 12, a relationship processing module 13, a first loss generation module 14, a character acquisition module 15, a phrase prediction module 16, a second A loss generation module 17 , a trigger detection generation module 18 , a sequence labeling module 19 , a third loss generation module 20 and a model training module 21 .
  • the model input module 11 and the semantic information extraction module 12 are used to extract the semantic features of the sample dialogue text and the sample argument pairs through the initial relationship prediction model to obtain the sample text semantic information, and each sample argument in the sample argument pair Belongs to said sample dialogue text;
  • the relationship processing module 13 and the first loss generation module 14 are used to perform argument relationship prediction based on the sample text semantic information, and determine the first loss based on the argument relationship prediction result, and the argument relationship prediction result is used to characterize the Describe the relationship between sample arguments;
  • the character acquisition module 15, the phrase prediction module 16 and the second loss generation module 17 are used to perform hidden character prediction based on the sample text semantic information, and determine the second loss based on the hidden character prediction result, and the hidden character prediction result is used for characterizing hidden characters in said sample dialogue text;
  • the trigger detection generation module 18, the sequence labeling module 19 and the third loss generation module 20 are used to perform trigger word prediction based on the semantic information of the sample text and the actual argument relationship, and determine the third loss based on the trigger word prediction result, the
  • the actual argument relationship is the labeled argument relationship corresponding to the sample argument, and the trigger word prediction result is used to represent the position of the trigger word in the sample dialogue text;
  • a model training module 21 configured to train the initial relationship prediction model based on the first loss, the second loss, and the third loss to obtain a dialogue relationship prediction model.
  • model input module 11 and the semantic information extraction module 12 are further configured to input the sample dialogue text and the sample argument pairs into the initial relationship prediction model, and in the initial relationship In the prediction model, semantic feature extraction is performed on the sample dialogue text and the sample argument pair to obtain the semantic information of the sample text;
  • a relationship processing module 13 configured to predict the predicted argument relationship between the sample arguments based on the sample text semantic information
  • a first loss generating module 14 configured to generate the first loss according to the actual argument relationship and the predicted argument relationship between the sample arguments
  • a character acquisition module 15 configured to acquire hidden characters in the sample dialogue text
  • Phrase prediction module 16 for predicting the predicted character corresponding to the hidden character based on the sample text semantic information
  • the second loss generation module 17 is used to generate a second loss according to hidden characters and predicted characters;
  • Trigger detection generating module 18 for generating trigger word detection text data according to actual argument relationship and sample text semantic information
  • Sequence labeling module 19 for predicting the corresponding predictive sequence labeling of trigger word detection text data
  • the third loss generation module 20 is used to generate the third loss according to the actual sequence label and the predicted sequence label corresponding to the trigger word detection text data;
  • the model training module 21 is used to adjust the model parameters of the initial relationship prediction model according to the first loss, the second loss and the third loss to generate a dialogue relationship prediction model; the dialogue relationship prediction model is used to predict the target argument in the target dialogue text The target argument relationship between the pairs.
  • phrase prediction module 16 includes:
  • the first determining unit 161 is configured to determine the mask hidden state corresponding to the hidden character from the sample text semantic information, and the mask hidden state is used to characterize the corresponding hidden character in the sample dialogue text semantic information;
  • the character predicting unit 162 is configured to predict the predicted character corresponding to the hidden character based on the hidden state of the mask.
  • the device also includes:
  • a sample acquisition module configured to acquire the original dialogue text and the sample argument pair
  • a hidden character determination module configured to determine the hidden character based on the sample argument pair in response to the argument type of at least one sample argument in the sample argument pair being a speaker argument;
  • a concealment processing module configured to perform concealment processing on the original dialogue text based on the hidden characters to obtain the sample dialogue text.
  • the trigger detection generating module 18 includes:
  • the second determination unit 181 is configured to determine an argument relationship vector corresponding to the actual argument relationship
  • the generating unit 182 is configured to splice the argument relationship vector and the semantic information of the sample text to generate the trigger word detection text data.
  • the generating unit 182 is also configured to:
  • the trigger word detection text corresponding to each sample dialogue phrase is determined as the trigger word detection text data
  • the sequence labeling module 19 includes:
  • the trigger word prediction unit 191 is configured to perform trigger word prediction based on the trigger word detection text data, and obtain the predicted phrase annotations corresponding to each of the sample dialogue phrases, and the predicted phrase annotations are used to represent the triggers to which the sample dialogue phrases belong. word type;
  • the third determining unit 192 is configured to mark the predicted phrase corresponding to each of the sample dialogue phrases as the predicted sequence label.
  • the semantic information extraction module 12 includes:
  • a sample splicing unit 121 configured to splice sample dialogue text and sample argument pairs based on sample text splicing symbols in the initial relationship prediction model to generate sample spliced text data
  • the argument replacement unit 122 is used to replace the first sample argument in the sample spliced text data with the first sample argument symbol, and replace the second sample argument in the sample spliced text data with the second sample argument symbol, generating sample text sequence data, the sample argument pair comprising the first sample argument and the second sample argument;
  • the information extraction unit 123 is configured to perform semantic feature extraction on the sample text sequence data to obtain sample text semantic information corresponding to the sample dialogue text.
  • sample text splicing symbols include sample global semantic symbols;
  • sample text sequence data includes sample dialogue sequence data corresponding to the sample dialogue text, and the sample dialogue sequence data includes N sample dialogue phrases, where N is a positive integer;
  • the information extraction unit 123 includes:
  • the hidden layer extraction subunit 1231 is used to perform hidden layer feature extraction on the sample global semantic symbols, N sample dialogue phrases, the first sample argument symbol and the second sample argument symbol in the sample text sequence data to obtain the sample
  • the semantic determination subunit 1232 is used to determine the sample global hidden state, N sample phrase hidden states, the first initial sample argument hidden state and the second initial sample argument hidden state as the sample text semantic information corresponding to the sample dialogue text .
  • the hidden layer extraction subunit 1231 includes:
  • the global processing subunit 123a is used to obtain the sample global relationship between the N sample dialogue phrases, the first sample argument symbol and the second sample argument symbol and the sample global semantic symbol respectively, and perform feature fusion on the sample global relationship, Generate sample global hidden states corresponding to sample global semantics;
  • the state extraction subunit 123b is used to perform hidden layer feature extraction on N sample dialogue phrases, the first sample argument symbol and the second sample argument symbol respectively, and obtain the sample phrase hidden states corresponding to the N sample dialogue phrases, The first initial sample argument hidden state corresponding to the first sample argument symbol, and the second initial sample argument hidden state corresponding to the second sample argument symbol.
  • the N sample dialogue phrases include the first sample argument symbol and the second sample argument symbol
  • the relationship processing module 13 includes:
  • the first argument identification unit 131 is configured to obtain at least one first sample phrase hidden state corresponding to the first sample argument symbol from the N sample phrase hidden states included in the sample text semantic information;
  • the second argument recognition unit 132 is used to obtain at least one second sample phrase hidden state corresponding to the second sample argument symbol from the N sample phrase hidden states;
  • the hidden state splicing unit 133 is used to splice the global hidden state of the sample, the hidden state of the first sample argument and the hidden state of the second sample argument to obtain the hidden state information of the sample;
  • the relationship predicting unit 134 is configured to predict the predicted argument relationship between the first sample argument and the second sample argument based on the hidden state information of the sample.
  • the relationship prediction unit 134 includes:
  • Semantic enhancement subunit 1341 configured to perform semantic enhancement on sample hidden state information to obtain sample enhanced semantic information
  • the probability selection subunit 1342 is used to determine the sample relationship prediction probability of M candidate argument relationships corresponding to the first sample argument and the second sample argument based on the sample enhanced semantic information, M is a positive integer;
  • the candidate argument relationship corresponding to the maximum sample relationship prediction probability is determined as the predicted argument relationship between the first sample argument and the second sample argument.
  • the embodiment of the present application provides a dialog relationship processing device.
  • the initial phrase prediction network and the initial trigger word prediction network By adding the initial phrase prediction network and the initial trigger word prediction network to the initial relationship prediction model, the hidden speaker arguments in the sample dialogue text are processed by the initial phrase prediction network. Prediction, and the trigger word prediction network predicts the trigger words that can guide the argument relationship in the sample dialogue text, so as to introduce additional second loss and third loss in the loss, and provide auxiliary information for the argument relationship prediction: speaker Features and trigger word features enable the initial relationship prediction model to learn more effective information that is more conducive to predicting the argument relationship, thereby improving the prediction accuracy of the argument relationship.
  • FIG. 10 is a schematic diagram of another dialog relationship processing device provided by an embodiment of the present application.
  • the dialog relationship processing device can be a computer program (including program code, etc.) running in a computer device, for example, the dialog relationship processing device can be an application software; the device can be used to execute the method in the embodiment of the present application corresponding steps.
  • the dialog relationship processing apparatus 1000 may include: an object model input module 31 , an object semantic acquisition module 32 and an object relationship prediction module 33 .
  • the target model input module 31 is used to input the target dialogue text and the target argument into the dialogue relationship prediction model
  • the target semantic acquisition module 32 is used to extract the semantic features of the target dialogue text and the target argument pair in the dialogue relationship prediction model, and obtain the target text semantic information corresponding to the target dialogue text, and each target theory in the target argument pair element belongs to said target dialogue text;
  • the target relationship prediction module 33 is used to predict the argument relationship of the target text semantic information based on the dialogue relationship prediction model, and obtain the target argument relationship between the target arguments; the dialogue relationship prediction model adopts the dialogue relationship as described in the above aspect obtained by processing method training.
  • the target semantic acquisition module 32 includes:
  • the target splicing unit 321 is used to splice the target dialog text and the target argument based on the target text splicing symbol to generate target spliced text data;
  • the target replacement unit 322 is used to replace the first target argument in the target spliced text data with the first target argument symbol, replace the second target argument in the target spliced text data with the second target argument symbol, and generate target text sequence data, said target argument pair comprising said first target argument and said second target argument;
  • the target acquisition unit 323 is configured to perform semantic feature extraction on the target text sequence data to obtain target text semantic information corresponding to the target dialogue text.
  • the target model input module 31 includes:
  • a consultation acquiring unit 311, configured to acquire the dialogue consultation information associated with the target dialogue text and the target dialogue text;
  • the dialogue analysis unit 312 is configured to analyze the dialogue consultation information, and extract the target argument pair indicated by the dialogue consultation information;
  • a model input unit 313, configured to input the target dialogue text and the target argument pair into the dialogue relationship prediction model
  • the device 900 also includes:
  • the data storage module 34 is configured to generate target question and answer data according to the target dialogue text, dialogue consultation information and target argument pairs, and add the target question and answer data to the question and answer database.
  • FIG. 11 is a schematic structural diagram of a computer device provided by an exemplary embodiment of the present application.
  • the computer device in this embodiment of the present application may include: one or more processors 1101 , a memory 1102 and an input/output interface 1103 .
  • the processor 1101 , memory 1102 and input/output interface 1103 are connected through a bus 1104 .
  • the memory 1102 is used to store computer programs, the computer programs include program instructions, and the input and output interface 1103 is used to receive data and output data, such as for data interaction between various networks in the model, or between computer equipment and user equipment Perform data interaction; the processor 1101 is used to execute the program instructions stored in the memory 1102 .
  • processor 1101 when the processor 1101 is located in a computer device for model training, it can perform the following operations:
  • the pair of sample arguments includes a first sample argument and a second sample argument
  • the predicted argument relationship between the first sample argument and the second sample argument is predicted, and the actual argument relationship and the predicted argument between the first sample argument and the second sample argument are predicted
  • the relation generates the first loss function
  • Generate trigger word detection text data according to the actual argument relationship and sample text semantic information, predict the predicted sequence annotation in the trigger word detection text data, and generate a third loss function according to the actual sequence annotation and predicted sequence annotation in the trigger word detection text data;
  • the dialogue relationship prediction model is used to predict the relationship between target argument pairs in the target dialogue text target argument relationship.
  • processor 1101 when the processor 1101 is located in a computer device for model prediction, it can perform the following operations:
  • the target argument pair includes a first target argument and a second target argument
  • the semantic information of the target text is predicted, and the target argument relationship between the first target argument and the second target argument is obtained.
  • the processor 1101 may be a central processing unit (Central Processing Unit, CPU), and the processor may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), dedicated integrated Circuit (Application Specific Integrated Circuit, ASIC), off-the-shelf programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • a general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.
  • the memory 1102 may include read-only memory and random-access memory, and provides instructions and data to the processor 1101 and the input-output interface 1103 .
  • a portion of memory 1102 may also include non-volatile random access memory.
  • memory 1102 may also store device type information.
  • the computer device can execute the implementation manners provided by the above-mentioned various method embodiments through various built-in functional modules.
  • the embodiment of the present application provides a computer device, including: a processor, an input and output interface, and a memory.
  • the processor obtains the computer program in the memory, executes each step of the method shown in FIG. 4, and performs dialogue relationship processing operations.
  • the initial phrase prediction network predicts the hidden speaker arguments in the sample dialogue text
  • the trigger word prediction network predicts
  • the trigger words that can guide the argument relationship are predicted, so as to introduce additional second loss and third loss in the loss, and provide auxiliary information for the argument relationship prediction: speaker features and trigger word features, so that the initial relationship
  • the predictive model can learn effective information that is more conducive to predicting the relationship between arguments, thereby improving the prediction accuracy of the relationship between arguments.
  • the embodiment of the present application also provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by the processor and executing the dialog relationship provided by each step in Figure 4 or Figure 8
  • the processing method refer to the implementation manner provided by each step in FIG. 4 or FIG. 8 for details, and details are not repeated here.
  • the description of the beneficial effect of adopting the same method will not be repeated here.
  • a computer program can be deployed to be executed on one computer device, or on multiple computer devices at one site, or distributed across multiple sites and interconnected by a communication network implement.
  • the computer-readable storage medium may be the dialog relationship processing apparatus provided in any of the foregoing embodiments or an internal storage unit of the computer device, such as a hard disk or memory of the computer device.
  • the computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, SMC), a secure digital (secure digital, SD) card, Flash card (flash card), etc.
  • the computer-readable storage medium may also include both an internal storage unit of the computer device and an external storage device.
  • the computer-readable storage medium is used to store the computer program and other programs and data required by the computer device.
  • the computer-readable storage medium can also be used to temporarily store data that has been output or will be output.
  • the embodiment of the present application also provides a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in the various optional modes in FIG. 4 or FIG. 8 , realizing the Conversational relationship extraction is divided into three related sub-tasks, which are relationship prediction task, phrase prediction task and trigger word prediction task. By combining these three sub-tasks, the model can be jointly trained to make full use of the phrase prediction task and trigger word prediction task.
  • the effective information learned in the word prediction task and based on this effective information affects the relationship prediction task, thereby improving the accuracy of dialogue relationship processing.
  • each flow and/or of the method flow charts and/or structural diagrams can be implemented by computer program instructions or blocks, and combinations of processes and/or blocks in flowcharts and/or block diagrams.
  • These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a A device for realizing the functions specified in one or more steps of the flowchart and/or one or more blocks of the structural diagram.
  • These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions
  • the device implements the functions specified in one or more blocks of the flowchart and/or one or more blocks of the structural schematic diagram.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby
  • the instructions provide steps for implementing the functions specified in one or more steps of the flowchart and/or one or more blocks in the structural illustration.
  • the modules in the device of the embodiment of the present application can be combined, divided and deleted according to actual needs.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Machine Translation (AREA)

Abstract

提供了一种对话关系处理方法、计算机及可读存储介质，涉及计算机技术领域。对话关系处理方法包括：通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息；基于样本文本语义信息进行论元关系预测，并基于论元关系预测结果确定第一损失；基于样本文本语义信息进行隐藏字符预测，并基于隐藏字符预测结果确定第二损失；基于样本文本语义信息和实际论元关系进行触发词预测，并基于触发词预测结果确定第三损失；基于第一损失、第二损失和第三损失，训练初始关系预测模型，得到对话关系预测模型。采用对话关系处理方法，可以提高论元关系的预测准确性。

Description

一种对话关系处理方法、计算机及可读存储介质
本申请要求于2021年06月17日提交的申请号为2021106744763、发明名称为“一种对话关系处理方法、计算机及可读存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，尤其涉及一种对话关系处理方法、计算机及可读存储介质。
背景技术
目前在多种应用场景中均具有对对话式数据进行处理的需求，在数据处理过程中会涉及到对对话式数据中论元的分析过程，比如，论元关系抽取任务即用于确定对话式数据中任意论元之间的论元关系。
在论元关系抽取任务中，一般会将对话文本与论元对进行语义特征提取，得到相关语义特征表示，进而基于该语义特征表示预测论元对对应的论元关系。
在上述论元关系预测过程中，由于对话关系的语义特征表示抽取较为复杂，难以准确找到与论元相关的上下文信息，从而影响论元关系的预测准确性。
发明内容
本申请实施例提供了一种对话关系处理方法、计算机及可读存储介质，可以提高对话场景中的论元关系预测准确性。
一方面，本申请实施例提供了一种对话关系处理方法，所述方法应用于计算机设备，该方法包括：
通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息，所述样本论元对中的各个样本论元属于所述样本对话文本；
基于所述样本文本语义信息进行论元关系预测，并基于论元关系预测结果确定第一损失，所述论元关系预测结果用于表征所述样本论元之间的关系；
基于所述样本文本语义信息进行隐藏字符预测，并基于隐藏字符预测结果确定第二损失，所述隐藏字符预测结果用于表征所述样本对话文本中的隐藏字符；
基于所述样本文本语义信息和实际论元关系进行触发词预测，并基于触发词预测结果确定第三损失，所述实际论元关系为所述样本论元对应的标注论元关系，所述触发词预测结果用于表征触发词在所述样本对话文本中的位置；
基于所述第一损失、所述第二损失和所述第三损失，训练所述初始关系预测模型，得到对话关系预测模型。
另一方面，本申请实施例提供了一种对话关系处理方法，所述方法应用于计算机设备，该方法包括：
将目标对话文本及目标论元对输入对话关系预测模型，在对话关系预测模型中对目标对话文本及目标论元对进行语义特征提取，得到目标对话文本对应的目标文本语义信息，所述目标论元对中的各个目标论元属于所述目标对话文本；
基于对话关系预测模型对目标文本语义信息进行论元关系预测，得到目标论元之间的目标论元关系；所述对话关系预测模型是采用上述方面所述的对话关系处理方法训练得到的。
另一方面，本申请实施例提供了一种对话关系处理装置，该装置包括：
模型输入模块和语义信息提取模块，用于通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息，所述样本论元对中的各个样本论元属于 所述样本对话文本；
关系处理模块和第一损失生成模块，用于基于所述样本文本语义信息进行论元关系预测，并基于论元关系预测结果确定第一损失，所述论元关系预测结果用于表征所述样本论元之间的关系；
字符获取模块、词组预测模块以及第二损失生成模块，用于基于所述样本文本语义信息进行隐藏字符预测，并基于隐藏字符预测结果确定第二损失，所述隐藏字符预测结果用于表征所述样本对话文本中的隐藏字符；
触发检测生成模块、序列标注模块以及第三损失生成模块，用于基于所述样本文本语义信息和实际论元关系进行触发词预测，并基于触发词预测结果确定第三损失，所述实际论元关系为所述样本论元对应的标注论元关系，所述触发词预测结果用于表征触发词在所述样本对话文本中的位置；
模型训练模块，用于基于所述第一损失、所述第二损失和所述第三损失，训练所述初始关系预测模型，得到对话关系预测模型。
另一方面，本申请实施例提供了一种对话关系处理装置，该装置包括：
目标模型输入模块，用于将目标对话文本及目标论元对输入对话关系预测模型；
目标语义获取模块，用于在对话关系预测模型中对目标对话文本及目标论元对进行语义特征提取，得到目标对话文本对应的目标文本语义信息，所述目标论元对中的各个目标论元属于所述目标对话文本；
目标关系预测模块，用于基于对话关系预测模型中的关系预测网络对目标文本语义信息进行论元关系预测，得到目标论元之间的目标论元关系，所述对话关系预测模型是采用上述方面所述的对话关系处理方法训练得到的。
另一方面，本申请实施例提供了一种计算机设备，包括处理器、存储器、输入输出接口；
处理器分别与存储器和输入输出接口相连，其中，输入输出接口用于接收数据及输出数据，存储器用于存储计算机程序，处理器用于调用该计算机程序，以使包含该处理器的计算机设备执行本申请实施例一方面中的对话关系处理方法。
另一方面，本申请实施例提供了一种计算机可读存储介质，计算机可读存储介质存储有计算机程序，该计算机程序适于由处理器加载并执行，以使得具有该处理器的计算机设备执行本申请实施例一方面中的对话关系处理方法。
另一方面，本申请实施例提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述各种可选实现方式中提供的对话关系处理方法。
实施本申请实施例，将具有如下有益效果：
在本申请实施例中，通过在初始关系预测模型中增加初始词组预测网络以及初始触发词预测网络，由初始词组预测网络对样本对话文本中被隐藏的说话者论元进行预测，以及由触发词预测网络对样本对话文本中可以指导论元关系的触发词进行预测，以便在损失中引入额外的第二损失和第三损失，对论元关系预测提供辅助信息：说话者特征以及触发词特征，使得初始关系预测模型可以学习到更有利于预测论元关系的有效信息，进而提高论元关系的预测准确性。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例提供的一种对话关系处理的网络交互架构图；
图2是本申请实施例提供的一种模型训练场景示意图；
图3是本申请实施例提供的一种模型结构示意图；
图4示出了本申请一个示例性实施例示出的对话关系处理方法的流程图；
图5是本申请另一个示例性实施例提供的对话关系处理方法的流程图；
图6是本申请一个示例性实施例提供的初始关系预测模型的模型架构图；
图7是本申请一个示例性实施例提供的初始关系预测模型的原理示意图；
图8是本申请一个示例性实施例提供的对话关系处理方法的流程图；
图9是本申请一个示例性实施例提供的对话关系处理装置示意图；
图10是本申请实施例提供的另一种对话关系处理装置示意图；
图11是本申请实施例提供的一种计算机设备的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
在本申请实施例中，请参见图1，图1是本申请实施例提供的一种对话关系处理的网络交互架构图，本申请实施例可以由计算机设备101实现，该计算机设备101可以与用户设备(如用户设备102a、用户设备102b及用户设备102c等)之间进行数据交互。其中，若计算机设备101为用于进行模型训练的设备，则该计算机设备101可以从各个用户设备中获取样本对话文本及样本对话文本中的样本论元对，基于样本对话文本及样本论元对实现对模型的训练，得到用于预测论元对关系的对话关系预测模型。可选的，该样本对话文本可以是从各个用户设备中获取到的，也可以是直接从计算机设备101中获取(即存储于计算机设备101中)，或者可以基于云存储技术进行获取，也可以基于区块链网络进行获取，在此不做限制。若计算机设备101为用于进行模型预测的设备，即计算机设备101中运行有预先训练完成的对话关系预测模型，则该计算机设备101可以响应任意一个用户设备发送的关系预测请求，例如，响应用户设备102a发送的关系预测请求，获取该关系预测请求所包括的目标对话文本及目标论元对，基于对话关系预测模型对目标对话文本与目标论元对进行预测，得到目标论元对的目标论元关系。可选的，计算机设备101也可以是获取到存储于计算机设备101中的目标对话文本及目标论元对，基于对话关系预测模型对目标对话文本与目标论元对进行预测，得到目标论元对的目标论元关系等。换句话说，本申请中所涉及的数据可以是由用户设备提供的，也可以存储于计算机设备101中，或者可以基于云存储技术进行存储，或者可以存储于区块链网络中，在此不做限制。
为了提高对话关系预测模型的关系预测准确性，本申请实施例在对话关系预测模型的训练任务中，额外增加了说话者预测任务以及触发词预测任务，其中，说话者预测任务用于预测样本对话文本中被遮盖的说话者论元，而触发词预测任务用于预测样本对话文本中表征论元关系的词语，这两个任务均可以辅助对话关系预测任务。
进一步地，可以参见图2，图2是本申请实施例提供的一种模型训练场景示意图。如图2所示，计算机设备可以获取模型训练样本201，该模型训练样本201包括样本对话文本及样本对话文本中的样本论元对，将样本对话文本及样本论元对输入初始关系预测模型，在初始关系预测模型中对样本对话文本及样本论元对进行语义特征提取，得到样本对话文本对应的样本文本语义信息，将样本文本语义信息输入初始关系预测模型中的任务处理区域，包括关系预测任务区域2021、词组预测任务区域2022及触发词预测任务区域2023等。其中，该样本论元对包括第一样本论元及第二样本论元。其中，该关系预测任务区域2021用于对样本论 元对进行关系抽取，该关系抽取是一种从自然语言文本中识别并判定实体对之间存在的特定关系的技术。具体的，在关系预测任务区域2021中，基于样本文本语义信息进行论元关系预测，得到第一样本论元及第二样本论元之间的预测论元关系，获取第一样本论元与第二样本论元之间的实际论元关系，根据预测论元关系与实际论元关系生成第一损失。在词组预测任务区域2022中，从样本文本语义信息中获取样本对话文本中隐藏字符所对应的掩码隐状态，进而基于该掩码隐状态预测隐藏字符对应的预测字符，进而根据隐藏字符与预测字符生成第二损失。在触发词预测任务区域2023中，根据实际论元关系与样本文本语义信息生成触发词检测文本数据，预测该触发词检测文本数据中的预测序列标注，预测标注序列用于指示样本对话文本中各个样本对话词组所属的触发词类型；根据触发词检测文本数据中的实际序列标注与预测序列标注生成第三损失。根据第一损失、第二损失及第三损失，对初始关系预测模型进行多任务学习，生成对话关系预测模型。通过在关系预测任务中引入词组预测任务及触发词预测任务等学到的有效信息，使得该关系预测任务中可以更为全面地获取到对话文本(如样本对话文本或目标对话文本等)中的信息，从而提高该关系预测任务的预测准确性。
进一步地，可以参见图3，图3是本申请实施例提供的一种模型结构示意图。如图3所示，初始关系预测模型中包括初始关系预测网络、初始词组预测网络及初始触发词预测网络等，该初始关系预测模型中还包括初始语言表征网络，该初始关系预测网络、初始词组预测网络及初始触发词预测网络等的输入，是基于该初始语言表征网络的输出得到的。在对该初始关系预测模型进行模型参数调整(即对初始关系预测模型的模型参数进行调整)后，生成对话关系预测模型，其中，该模型参数是使用通用变量来建立函数和变量之间关系的一个数量，在人工智能领域中，模型参数通常是实数矩阵，可选的，随着模型的发展，该模型参数也可以是其他格式的数据。其中，该对话关系预测模型可以包括语言表征网络、关系预测网络、词组预测网络及触发词预测网络。可选的，该对话关系预测模型也可以只包括语言表征网络及关系预测网络等，以精简模型结构。
其中，该词组预测网络及触发词预测网络用于对关系预测网络进行特征信息补充，从而提高关系预测网络的预测准确性。其中，关系预测任务区域是指初始关系预测网络及关系预测网络所在的区域，也就是说，关系预测任务区域中的网络在训练前称为初始关系预测网络，在训练后称为关系预测网络，用于预测论元对(如样本论元对或目标论元对等)之间的论元关系；词组预测任务区域是指初始词组预测网络及词组预测网络所在的区域，也就是说，词组预测任务区域中的网络在训练前称为初始词组预测网络，在训练后称为词组预测网络，用于对对话文本中的隐藏词组进行预测；触发词预测任务区域是指初始触发词预测网络及触发词预测网络所在的区域，也就是说，触发词预测任务区域中的网络在训练前称为初始触发词预测网络，在训练后称为触发词预测网络，用于预测对话文本中各个对话词组是否属于触发词。
可以理解的是，本申请实施例中所提及的计算机设备或用户设备包括但不限于终端设备或服务器。换句话说，计算机设备或用户设备可以是服务器或终端设备，也可以是服务器和终端设备组成的系统。其中，以上所提及的终端设备可以是一种电子设备，包括但不限于手机、平板电脑、台式电脑、笔记本电脑、掌上电脑、车载设备、增强现实/虚拟现实(Augmented Reality/Virtual Reality，AR/VR)设备、头盔显示器、智能电视、可穿戴设备、智能音箱、数码相机、摄像头及其他具备网络接入能力的移动互联网设备(Mobile Internet Device，MID)，或者火车、轮船、飞行等场景下的终端设备等。其中，以上所提及的服务器可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、车路协同、内容分发网络(Content Delivery Network，CDN)、以及大数据和人工智能平台等基础云计算服务的云服务器。
请参考图4，其示出了本申请一个示例性实施例示出的对话关系处理方法的流程图，本 申请实施例以该方法应用于图1所示的计算机设备为例进行示例性说明，该方法包括：
步骤401，通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息，样本论元对中的各个样本论元属于样本对话文本。
其中，初始关系预测模型包括初始语言表征网络、初始关系预测网络、初始词组预测网络以及初始触发词预测网络，其中，初始语言表征网络用于提取样本文本语义信息(也即文本语义特征表示)，而初始关系预测网络、初始词组预测网络以及初始触发词预测网络共享该样本文本语义信息，由初始关系预测网络基于样本文本语义信息预测论元关系，初始词组预测网络基于样本文本语义信息，预测样本对话文本中的隐藏字符，初始触发词预测网络基于样本文本语义信息，预测样本对话文本中各个样本词组是否为触发词。可选的，初始关系预测模型中还包括语义增强网络，语义增强网络用于对提取到的样本文本语义信息进行语义增强。
示意性的，初始语言表征网络可以使用双向特征表示的自编码预训练语言模型(Bidirectional Encoder Representations from Transformers，BERT)、自回归预训练模型(Embedding from Language Model，ELMo)或生成式的预训练模型(Generative Pre-Training，GPT/GPT2)等，本申请实施例对初始语言表征网络所采用的网络类型不构成限定。
基于上述初始关系预测模型中各个功能模块的作用，在一种可能的实施方式中，计算机设备将样本对话文本和样本论元对输入初始关系预测模型中，由初始关系预测模型中的初始关系预测网络，对样本对话文本和样本论元对进行语义特征提取，以得到样本文本语义信息，而样本文本语义信息即样本对话文本以及样本论元对相对应的文本语义特征表示，也即样本文本语义信息为特征向量的形式。
在训练初始关系预测模型之前，需要预先准备训练样本集，本申请实施例中的训练样本集为：样本对话文本和样本论元对，其中，样本论元对中的各个样本论元属于样本对话文本，也即模型训练目标为：如何使得初始关系预测模型可以基于样本对话文本，准确预测出样本论元之间的论元关系。
其中，样本对话文本可以是一段经过m轮对话过程中所生成的对话语句组成，示例性的，样本对话文本可以表示为：d＝s 1：t 1，s 2：t 2，…，s m：t m；其中，s 1表示第一个说话者，t 1表示第一个说话者对应的说话内容，s m表示第m个说话者，t m表示第m个说话者对应的说话内容，t m可以由N个样本对话词组(词语)构成。
其中，样本论元对中的各个样本论元属于样本对话文本，一般来说样本论元对中包含第一样本论元和第二样本论元，也即第一样本论元和第二样本论元属于样本对话文本，则初始关系预测模型用于预测第一样本论元和第二样本论元之间的论元关系；可选的，样本论元对也可以包含两个以上的样本论元，则初始关系预测模型用于预测样本论元对中任意两个样本论元之间的论元关系。
可选的，样本论元对中各个样本论元的论元类型可以都是说话者论元，示例性的，以样本对话文本是d＝s 1：t 1，s 2：t 2，…，s m：t m为例，样本论元对可以包括：s 1和s 3，样本论元对也可以包括：s 1和s 2；或，样本论元对中各个样本论元的论元类型可以既包含说话者论元，也包含词组论元，词组论元为样本对话文本中除说话者论元之外的任意论元，示例性的，样本论元对可以是s 1和N 1，N 1可以表示说话内容(t)中包含的任意词语；或，样本论元对中各个样本论元的论元类型均为词组论元，也即样本论元对中不包含说话者论元，示例性的，样本论元对可以是：N 1和N 2，N 1和N 2可以表示说话者内容(t)中包含的任意两个不同词语。
基于上文中样本论元对与样本对话文本的关系，在一种可能的实施方式中，人工将样本对话文本和样本论元对输入计算机设备中，用于后续训练初始关系预测模型；可选的，也可以人工在样本对话文本中标注出样本论元对，将标注有样本论元对的样本对话文本(标注文本)输入计算机设备，由计算机设备从该标注文本中确定出样本对话文本以及样本论元对；可选的，计算机设备中设置有样本论元对确定方式，将样本对话文本输入计算机设备中，由 计算机设备基于输入的样本对话文本，以及样本论元对确定方式，从样本对话文本中提取出样本论元对。
其中，样本论元对确定方式可以指示各个样本论元的论元类型，比如，从样本对话文本中提取论元类型为说话者论元的样本论元，则对应将任意两个说话者论元确定为样本论元对。
步骤402，基于样本文本语义信息进行论元关系预测，并基于论元关系预测结果确定第一损失，论元关系预测结果用于表征样本论元之间的关系。
由步骤401得到的样本文本语义信息会被三个任务共享，在一种可能的实施方式中，将样本文本语义信息输入初始关系预测网络中，初始关系预测网络从样本文本语义信息中，学习到样本对话文本的全局语义，以及每个论元的上下文相关语义，进而基于全局语义和上下文相关语义，预测样本论元之间的关系，进一步基于论元关系预测结果确定第一损失。
其中，若样本论元对包含第一样本论元和第二样本论元，则论元关系预测结果为第一样本论元和第二样本论元之间的预测论元关系，第一损失由预测论元关系和实际论元关系确定，实际论元关系为第一样本论元和第二样本论元之间的真实论元关系。
可选的，初始关系预测网络为一个多分类器，初始关系预测网络输出的论元关系预测结果为：样本论元对对应论元关系是各个候选论元关系的概率，并选择概率最大的候选论元关系，将其确定为样本论元对对应的预测论元关系。
示例性的，若样本论元对为两个说话者论元，则论元关系预测结果为两个说话者论元之间的关系，初始关系预测网络的输出可以为：P 1＝0.05(表示论元关系为母女关系的概率为0.05)、P 2＝0.01(表示论元关系为母子关系的概率为0.01)、P 3＝0.8(表示论元关系为兄妹关系的概率为0.8)、P 4＝0.05(表示论元关系为父女关系的概率为0.05)、P 5＝0.04(表示论元关系为父子关系的概率为0.05)，比较各个概率的大小关系，将兄妹关系确定为样本论元对对应的预测论元关系。
步骤403，基于样本文本语义信息进行隐藏字符预测，并基于隐藏字符预测结果确定第二损失，隐藏字符预测结果用于表征样本对话文本中的隐藏字符。
在对话关系预测场景中，考虑到说话者论元之间的论元关系抽取与说话者具有高度相关的特点，比如，说话者论元之间的关系与说话者的特征(表达习惯、说话语气等)息息相关，因此，为了进一步辅助论元关系预测任务，使得初始关系预测模型更好学习到与说话者相关的特征，在一种可能的实施方式中，部署有初始词组预测网络(也可以称之为初始说话者预测网络)，用于预测样本对话文本中的隐藏字符，也就是说，用于预测样本对话文本中被隐藏的说话者论元。
为了预测样本对话文本中的隐藏字符，则样本对话文本中需要存在被遮盖或隐藏的说话者论元，对应的，在计算机设备将样本对话文本输入初始关系预测模型之前，还需要对原始对话文本中说话者论元对应的说话者词语进行掩盖，得到样本对话文本，再进行后续语义特征提取过程。
在一种可能的实施方式中，样本对话文本中包含有隐藏字符，在基于样本对话文本以及样本论元对进行语义特征提取过程中，无论对于隐藏字符还是未隐藏字符，均可以得到其对应的语义特征表示，也即样本文本语义信息中包含有隐藏字符相对应的语义特征表示，则将该语义特征表示输入初始词组预测网络中进行隐藏字符预测，得到初始词组预测网络输出的隐藏字符预测结果，进而基于实际隐藏字符和隐藏字符预测结果，确定出第二损失。
可选的，由于初始词组预测网络主要是用于预测样本对话文本中被隐藏的说话者词语(说话者论元对应的词语)，任务原理为：如果初始词组预测网络可以分辨出是哪一个说话者说的话，表示初始词组预测网络是可以基于样本文本语义信息分析出说话者特征的，进一步表明样本文本语义信息提取较为准确，从而有助于预测与说话者相关的论元关系，也就是说，初始词组预测网络可以为预测说话者论元之间的论元关系提供辅助作用；换句话说，若初始关系预测模型无需预测说话者论元之间的论元关系，则无需部署初始词组预测网络；若初始关 系预测模型需要预测说话者论元之间的论元关系，则需要部署初始词组预测网络。
步骤404，基于样本文本语义信息和实际论元关系进行触发词预测，并基于触发词预测结果确定第三损失，实际论元关系为样本论元对应的标注论元关系，触发词预测结果用于表征触发词在样本对话文本中的位置。
在对话关系预测场景中，对话文本中往往存在可以指示论元关系的词语，称之为触发词，对于结构复杂且篇幅冗长的对话文本，触发词信息对于最终的论元关系抽取也可以起到一定的指导作用，因此，在一种可能的实施方式中，在初始关系预测模型中部署有初始触发词预测网络，用于预测触发词在样本对话文本中的位置，也即初始触发词预测网络需要识别样本对话文本中的各个样本对话词组是否为触发词。
示意性的，若样本对话文本为一对母女之间的对话，则触发词可能是样本对话文本中的出现的“妈妈”“女儿”等对话词组。
若需要识别样本对话词组是否为触发词，也就是需要判断该样本对话词组是否与论元关系存在某种联系，或者样本对话词组是否可以指示论元关系，因此，在模型训练过程中，若需要引导初始触发词预测网络正确学习到触发词特征，需要真实论元关系参与训练，在一种可能的实施方式中，计算机设备获取到样本论元对对应的实际论元关系，并将该实际论元关系和样本文本语义信息共同输入初始触发词预测网络，以便初始触发词预测网络可以基于样本文本语义信息，以及实际论元关系，预测各个样本对话词组是否为触发词，从而得到可以表征触发词在样本对话文本中位置的触发词预测结果，进而基于触发词预测结果和实际触发词信息，确定第三损失。
步骤405，基于第一损失、第二损失和第三损失，训练初始关系预测模型，得到对话关系预测模型。
在一种可能的实施方式中，对提出的关系抽取任务、说话者预测任务以及触发词预测任务，以多任务联合学习的方式进行联合训练，也即根据第一损失(论元关系预测损失)、第二损失(说话者预测损失)以及第三损失(触发词预测损失)，建立联合损失，进而基于联合损失训练初始关系预测模型，经过多轮重复训练直至损失收敛后，得到对话关系预测模型。
可选的，说话者预测任务可以从说话者特征角度，为关系预测任务提供辅助，而触发词预测任务可以从触发词特征角度，为关系预测任务提供指导，在其他可能的实施方式中，可以采用说话者预测任务和触发词预测任务中的至少一种，为关系预测任务提供指导，也即初始关系预测模型可以包括：初始语言表征网络、初始关系预测网络、初始词组预测网络以及初始触发词预测网络，对应的，在模型训练过程中，基于第一损失(由初始关系预测网络生成)、第二损失(由初始词组预测网络生成)以及第三损失(由初始触发词预测网络生成)共同训练初始关系预测模型；可选的，初始关系预测模型中还可以包括：初始语言表征网络、初始关系预测网络以及初始词组预测网络，对应的，在模型训练过程中，基于第一损失和第二损失共同训练初始关系预测模型；可选的，初始关系预测模型中还可以包括：初始语言表征网络、初始关系预测网络以及初始触发词预测网络，对应的，在模型训练过程中，基于第一损失和第三损失共同训练初始关系预测模型。
需要说明的是，在模型训练阶段会增加有初始触发词预测网络，以及初始词组预测网络来辅助训练，但是在模型应用阶段，可以无需上述两个子任务，仅保留关系预测任务以及语义特征提取任务，也即对话关系预测模型中可以仅包含语言表征网络以及关系预测网络。
综上所述，本申请实施例中，通过在初始关系预测模型中增加初始词组预测网络以及初始触发词预测网络，由初始词组预测网络对样本对话文本中被隐藏的说话者论元进行预测，以及由触发词预测网络对样本对话文本中可以指导论元关系的触发词进行预测，以便在损失中引入额外的第二损失和第三损失，对论元关系预测提供辅助信息：说话者特征以及触发词特征，以便初始关系预测模型可以学习到更有利于预测论元关系的有效信息，进而提高论元关系的预测准确性。
进一步地，请参见图5，图5是本申请另一个示例性实施例提供的对话关系处理方法的流程图。如图5所示，该对话关系处理方法包括如下步骤：
步骤501，将样本对话文本及样本论元对输入初始关系预测模型，在初始关系预测模型中对样本对话文本及样本论元对进行语义特征提取，得到样本文本语义信息。
其中，样本论元对包括第一样本论元及第二样本论元。
在本申请实施例中，计算机设备获取训练样本，该训练样本包括样本对话文本及样本对话文本中的样本论元对，其中，论元是指带有题元角色的体词性成分，跟谓词搭配的名词可以称为论元，该样本对话文本可以是包括至少两个论元的任意一个文本数据，该样本论元对可以是样本对话文本中的任意两个论元，例如，由样本对话文本中的任意两个说话者组成的论元对，或者，由样本对话文本中的任意两个物品组成的论元对，或者，也可以是由样本对话文本中的说话者与物品组成的论元对等，在此不做限制。举例来说，获取到样本对话文本为“Speaker 1：‘Uh，like，could these margaritas be any stronger？Hey，Chandler.’，Speaker 2：‘Hello，Mr.Bing’，Speaker 3：‘…’，…”，则该样本论元对可以是该样本对话文本中的任意两个说话者组成的论元对，如“{Speaker 1，Speaker 2}、{Speaker 1，Speaker 3}及{Speaker 2，Speaker 3}等”，也可以是说话者与物品组成的论元对，如“{Speaker 1，margaritas}、{Speaker 2，margaritas}等”，在此不做限制。也就是说，该样本论元对可以有不同的论元类型，包括但不限于同说话者论元、同物品论元及说话者与物品论元类型等。其中，在模型训练过程中，可以同时基于不同的论元类型对初始关系预测模型进行训练，得到可以检测不同论元类型的对话关系预测模型，以提高训练得到的对话关系预测模型的泛化性；或者，也可以基于不同的论元类型，分别对初始关系预测模型进行训练，得到多个对话关系预测模型，不同对话关系预测模型所检测的论元类型不同，以提高训练得到的对话关系预测模型的准确性。可选的，可以基于需要，综合考虑模型的泛化性及准确性等，确定一个模型所需检测论元类型的数量，在此不做限制。进一步地，计算机设备可以将获取到的样本对话文本及样本对话文本中的样本论元对输入初始关系预测模型，在初始关系预测模型中对样本对话文本及样本论元对进行语义特征提取，得到样本对话文本对应的样本文本语义信息。
其中，计算机设备可以在初始关系预测模型中，对样本对话文本及样本论元对进行数据预处理及隐层特征提取，得到样本对话文本对应的样本文本语义信息。其中，图6是本申请一个示例性实施例提供的初始关系预测模型的模型架构图，如图6所示，计算机设备获取到训练样本601，该训练样本601可以包括样本对话文本及样本对话文本中的样本论元对，将样本对话文本及样本论元对输入初始关系预测模型602，基于该初始关系预测模型602中的共享编码器对样本对话文本及样本论元对进行数据预处理以及隐层特征提取，得到该样本对话文本对应的样本文本语义信息，其中，共享编码器也即上文实施例中所示出的初始语言表征网络。该初始关系预测模型602可以执行关系预测任务、词组预测任务及触发词预测任务；可选的，在训练阶段，基于关系预测任务、词组预测任务以及触发词预测任务进行联合训练；而在测试阶段和应用阶段，可以仅使用共享编码器和关系预测任务，得到关系预测结果。
本实施例中，在基于样本对话文本以及样本论元对，进行语义特征提取之前，还需要对样本对话文本以及样本论元对进行数据预处理，以便后续语义特征提取过程，其中，数据预处理过程可以大致包括两个步骤：样本数据拼接过程以及样本符号替换过程。
可选的，初始关系预测模型中还可以包括数据预处理模块，用于对样本对话文本以及样本论元对进行数据预处理过程。
在一个示例性的例子中，步骤501还可以包括步骤501A～步骤501C。
步骤501A，在初始关系预测模型中基于样本文本拼接符对样本对话文本与样本论元对进行拼接，生成样本拼接文本数据。
由于需要将样本对话文本以及样本论元对一起输入初始关系预测模型中，对应的，需要将样本对话文本和样本论元对拼接为一组文本数据，在一种可能的实施方式中，计算机设备 获取到样本对话文本以及样本论元对之后，基于样本文本拼接符拼接样本对话文本以及样本论元对。
其中，样本文本拼接符可以包括样本全局语义符和样本分隔符，其中，样本全局语义符用于指代输入初始关系预测模型的完整样本文本数据(包含样本对话文本以及样本论元对)，样本分隔符用于分隔样本对话文本以及样本论元对。示意性的，样本全局语义符可以表示为[CLS]，样本分隔符可以表示为[SEP]；可选的，样本全局语义符和样本分隔符并不仅限定于采用上述符号，也可以采用其他特殊符号，本申请实施例对此不构成限定。
在一个示例性的例子中，若样本对话文本表示为：d＝s 1：t 1，s 2：t 2，s 3：t 3…，s m：t m，样本论元对包含第一样本论元a 1和第二样本论元a 2，样本全局语义符为[CLS]，样本分隔符为[SEP]，对应的，将样本对话文本和样本论元对进行拼接后，生成的样本拼接文本数据可以表示为：[CLS]s 1：t 1，s 2：t 2，s 3：t 3…，s m：t m[SEP]a 1[SEP]a 2[SEP]。
具体的，计算机设备可以在初始关系预测模型中基于样本文本拼接符对样本对话文本与样本论元对进行拼接，生成样本拼接文本数据，其中，该样本文本拼接符可以包括样本全局语义符及样本分隔符等，具体的，计算机设备可以对样本对话文本与样本论元对进行拼接，在拼接得到的拼接样本中插入样本文本拼接符，得到样本拼接文本数据，其中，该样本文本拼接符及该样本文本拼接符的插入方式，可以是基于初始关系预测模型中的初始语言表征网络所确定，即不同初始语言表征网络有不同的样本拼接规范，比如，若初始语言表征网络采用BERT，对应的样本拼接符包括[CLS]和[SEP]，拼接符插入方式如上文实施例所示；该样本文本拼接符可以是任意一种特殊字符或特殊字符串，在此不做限制，该样本全局语义符用于是指该样本对话文本及样本论元对的全局表示，样本分割符用于对样本对话文本与样本论元对进行分割。举例来说，将该样本对话文本记作Tx，假定该样本文本拼接符中的样本全局语义符为[CLS]，样本文本拼接符中的样本分隔符为[SEP]，基于样本全局语义符与样本分割符对样本对话文本及样本论元对进行拼接，其中，将第一样本论元记作a 1，将第二样本论元记作a 2，在对样本对话文本及样本论元对进行拼接后，得到样本拼接文本数据，该样本拼接文本数据可以是“[CLS]Tx[SEP]a 1[SEP]a 2[SEP]”，其中，假定该样本对话文本Tx为“s 1：t 1s 2：t 2…s u：t u”，此时，该样本拼接文本数据可以表示为“[CLS]s 1：t 1s 2：t 2…s u：t u[SEP]a 1[SEP]a 2[SEP]”；可选的，也可以进一步基于样本对话文本的文本组成结构在样本对话文本中插入样本分割符，得到样本拼接文本数据，如“[CLS]s 1：t 1[SEP]s 2：t 2[SEP]…[SEP]s u：t u[SEP]a 1[SEP]a 2[SEP]”，在此不做限制。
步骤501B，将样本拼接文本数据中的第一样本论元替换为第一样本论元符号，将样本拼接文本数据中的第二样本论元替换为第二样本论元符号，生成样本文本序列数据。
由于初始关系预测模型的训练任务是：预测第一样本论元和第二样本论元之间的论元关系，为了在输入的样本拼接文本数据中突出强调第一样本论元和第二样本论元，使得初始关系预测模型(主要是初始关系预测模型中的初始语言表征网络)更关注于样本拼接文本数据中的样本论元，在一种可能的实施方式中，采用特殊符号替换样本拼接文本数据中的样本论元，该特殊符号既可以标示出样本论元在样本对话文本中的位置，使得初始语言表征网络更好的提取出与该样本论元相关的上下文语义信息，同时该特殊符号也与样本论元本身的论元类型相关，使得初始语言表征网络可以提取出样本论元类型相关的特征，进一步提高语义特征提取过程中的信息丰富性。
可选的，特殊符号是专门用于表示说话者论元的，换句话说，当样本论元对中包含说话者类型的样本论元时，对样本拼接文本数据中的样本论元进行替换；在一种可能的实施方式中，若样本论元中的第一样本论元和第二样本论元均为说话者类型的论元，则将样本拼接文本数据中的第一样本论元替换为第一样本论元符号，将样本拼接文本数据中的第二样本论元替换为第二样本论元符号，从而生成样本文本序列数据。需要说明的是，由于样本拼接文本数据中包含样本对话文本部分以及样本论元对部分，则进行替换时，既需要对样本对话文本 部分所包含的第一样本论元和第二样本论元进行替换，也需要将样本论元对部分的第一样本论元和第二样本论元进行替换。
在一个示例性的例子中，样本拼接文本数据为：[CLS]s 1：t 1，s 2：t 2，s 3：t 3…，s m：t m[SEP]a 1[SEP]a 2[SEP]，且a 1和a 2均为说话者类型的样本论元，a 1表示s 1，a 2表示s 2，用于预测样本对话文本中说话者1和说话者2之间的关系；第一样本论元符号为S 1，第二样本论元符号为S 2，基于第一样本论元符号和第二样本论元符号对样本拼接文本数据进行替换，得到的样本文本序列数据可以为：[CLS]S 1：t 1，S 2：t 2，s 3：t 3…，s m：t m[SEP]S 1[SEP]S 2[SEP]。
可选的，在其他可能的实施方式中，为了进一步突出强调样本论元对在样本拼接文本数据中所处的位置，还可以采用起始论元符号以及结束论元符号来标注各个样本论元在样本拼接文本数据中的位置。其中，起始论元符号可以用[B]表示，结束论元符号可以用[E]表示；可选的，起始论元符号和结束论元符号还可以采用其他形式的符号，本申请实施例对此不构成限定。
在一个示例性的例子中，样本拼接文本数据为：[CLS]s 1：t 1，s 2：t 2，s 3：t 3…，s m：t m[SEP]a 1[SEP]a 2[SEP]，对其进行论元符号替换，以及增加起始论元符号和结束论元符号后，得到的样本文本序列数据可以为：[CLS][B]S 1[E]：t 1，[B]S 2[E]：t 2，s 3：t 3…，s m：t m[SEP]S 1[SEP]S 2[SEP]。
进一步地，计算机设备可以将样本拼接文本数据中的第一样本论元替换为第一样本论元符号，将样本拼接文本数据中的第二样本论元替换为第二样本论元符号，生成样本文本序列数据。可选的，在将样本论元替换为对应的样本论元符号(即将第一样本论元替换为第一样本论元符号，或将第二样本论元替换为第二样本论元符号)时，可以基于起始论元符号及结束论元符号标注样本论元在样本拼接文本数据中的位置。具体的，该样本文本序列数据的生成过程可以参见公式①所示：
Figure PCTCN2021108503-appb-000001
如公式①所示，
Figure PCTCN2021108503-appb-000002
表示样本文本序列数据中的第i个字符，[B]表示起始论元符号，[E]表示结束论元符号，a 1表示第一样本论元，a 2表示第二样本论元，s i表示样本拼接文本数据中的第i个字符，[S 1]表示第一样本论元符号，[S 2]表示第二样本论元符号，可选的，本申请汇总的第一样本论元符号、第二样本论元符号、起始论元符号及结束论元符号等均不限于上述的表示方式，也可以由其他的特殊字符或特殊字符串进行表示，在此不做限制。其中，公式①用于表示，计算机设备遍历样本拼接文本数据，若该样本拼接文本数据中的第i个字符为第一样本论元(即s i＝a 1)，则将样本拼接文本数据中的第i个字符替换为第一样本论元符号[S 1]，可选的，还可以基于起始论元符号[B]及结束论元符号[E]标注该第一样本论元在样本拼接文本数据中的位置，即，将样本拼接文本数据中的第i个字符替换为[B][S 1][E]；若该样本拼接文本数据中的第i个字符为第二样本论元(即s i＝a 2)，则将样本拼接文本数据中的第i个字符替换为第二样本论元符号[S 2]，可选的，还可以基于起始论元符号[B]及结束论元符号[E]标注该第二样本论元在样本拼接文本数据中的位置，即，将样本拼接文本数据中的第i个字符替换为[B][S 2][E]；若该样本拼接文本数据中的第i个字符不是第一样本论元，且不是第二样本论元，则将该样本拼接文本数据中的第i个字符保持不变。通过上述过程，得到样本文本序列数据。其中，i为正整数，i小于或等于样本拼接文本数据中所包括的字符数量。
需要说明的是，在生成样本拼接文本数据后，也可以无需对样本拼接文本数据进行样本符号替换，直接将样本拼接文本数据输入初始语言表征网络进行语义特征提取。
其中，上述基于样本对话文本及样本对话文本中的样本论元对生成样本文本序列数据的过程，可以认为是针对样本对话文本及样本对话文本中的样本论元对的数据预处理过程。
步骤501C，对样本文本序列数据进行语义特征提取，得到样本文本语义信息。
当对样本对话文本以及样本论元对进行数据预处理后，即可以将数据预处理后得到的样本文本序列数据输入初始语言表征网络，并由初始语言表征网络对样本文本序列数据进行语义特征提取，进而得到相关的样本文本语义信息。
由于样本文本序列数据中包含多种数据，比如，样本对话词组(样本对话文本中包含的词语)、样本论元对，则进行语义特征提取过程中，会分别针对不同类型的数据提取出与之相关的语义表示。
在一个示例性的例子中，对样本文本序列数据进行语义特征提取的过程可以包括步骤一和步骤二。
步骤一、对样本文本序列数据中的样本全局语义符、N个样本对话词组、第一样本论元符号及第二样本论元符号分别进行隐层特征提取，得到样本全局语义符对应的样本全局隐状态、N个样本对话词组分别对应的样本词组隐状态、第一样本论元符号对应的第一初始样本论元隐状态以及第二样本论元符号对应的第二初始样本论元隐状态。
其中，样本文本序列数据中包含样本全局语义符、第一样本论元符号、第二样本论元符号以及N个样本对话词组，样本对话词组为样本文本序列数据中组成样本对话文本的各个样本词语或样本字符。示例性的，若样本文本序列数据为：[CLS][B]S 1[E]：t 1，[B]S 2[E]：t 2，s 3：t 3…，s m：t m[SEP]S 1[SEP]S 2[SEP]，则样本对话词组为[B]S 1[E]：t 1，[B]S 2[E]：t 2，s 3：t 3…，s m：t m中所包含的词语或词组。语义特征提取的目的为：提取全局语义表示以及上下文语义表示，基于该语义特征提取任务，在一种可能的实施方式中，初始语言表征网络通过样本全局语义符识别输入的完整样本数据，对样本文本序列数据进行全局语义特征提取，得到样本全局隐状态(即全局语义表示)；并且，初始语言表征网络还对样本文本序列数据中的样本论元对进行上下文语义特征提取，得到与第一样本论元相对应的第一初始样本论元隐状态，以及与第二样本论元相对应的第二初始样本论元隐状态；同时，初始语言表征网络还对样本文本序列数据中的各个样本对话词组进行上下文语义特征提取，以得到与各个样本对话词组相对应的N个样本词组隐状态。
进一步地，计算机设备可以对样本文本序列数据进行特征提取(即隐层特征提取)，得到样本对话文本对应的样本文本语义信息。具体的，该样本文本拼接符包括样本全局语义符；样本文本序列数据包括样本对话文本对应的样本对话序列数据，样本对话序列数据包括N个样本对话词组，N为正整数。计算机设备在对样本文本序列数据进行特征提取，得到样本对话文本对应的样本文本语义信息时，具体是对样本文本序列数据中的样本全局语义符、N个样本对话词组、第一样本论元符号及第二样本论元符号分别进行隐层特征提取，得到样本全局语义符对应的样本全局隐状态、N个样本对话词组分别对应的样本词组隐状态、第一样本论元符号对应的第一初始样本论元隐状态以及第二样本论元符号对应的第二初始样本论元隐状态。将样本全局隐状态、N个样本词组隐状态、第一初始样本论元隐状态及第二初始样本论元隐状态，确定为样本对话文本对应的样本文本语义信息。
步骤二、将样本全局隐状态、N个样本词组隐状态、第一初始样本论元隐状态及第二初始样本论元隐状态，确定为样本对话文本对应的样本文本语义信息。
样本文本语义信息即包含全局语义表示和上下文语义表示，其中，全局语义表示为样本全局隐状态，而上下文语义表示为N个样本词组隐状态、第一初始样本论元隐状态以及第二初始样本论元隐状态；也即计算机设备将提取到的样本全局隐状态、N个样本词组隐状态、第一初始样本论元隐状态及第二初始样本论元隐状态，确定为样本文本语义信息。
具体的，计算机设备可以基于初始语言表征网络，获取N个样本对话词组、第一样本论元符号及第二样本论元符号分别与样本全局语义符之间的样本全局关系，对样本全局关系进行特征融合，生成样本全局语义符对应的样本全局隐状态；对N个样本对话词组、第一样本论元符号及第二样本论元符号分别进行隐层特征提取，得到N个样本对话词组分别对应的样本词组隐状态、第一样本论元符号对应的第一初始样本论元隐状态，以及第二样本论元符号 对应的第二初始样本论元隐状态。
举例来说，请参见图7，图7是本申请一个示例性实施例提供的初始关系预测模型的原理示意图。如图7所示，该初始关系预测模型包括语言表征任务701、关系预测任务702、词组预测任务703及触发词预测任务704。在样本文本语义信息提取阶段：计算机设备获取训练样本7011，该训练样本7011包括样本对话文本701a及样本论元对701b，若该样本对话文本701a为“a 1:Mom！a 2：Sweetie……”，该样本论元对701b包括第一样本论元a 1及第二样本论元a 2。对训练样本7011进行数据预处理(即进行样本数据拼接处理以及样本符号替换处理)，得到样本文本序列数据7012(本实施例中未显示起始论元符号以及结束论元符号)；其中，数据预处理过程为：基于样本文本拼接符对样本对话文本701a及样本论元对701b进行拼接，生成样本拼接文本数据，该样本拼接文本数据可以是“[CLS]a 1:Mom！a 2：Sweetie……[SEP]a 1[SEP]a 2[SEP]”。进一步地，将样本拼接文本数据中的第一样本论元a 1替换为第一样本论元符号S 1，将样本拼接文本数据中的第二样本论元a 2替换为第二样本论元符号S 2，得到样本文本序列数据7012，该样本文本序列数据7012为“[CLS]S 1:Mom！S 2：Sweetie……[SEP]S 1[SEP]S 2[SEP]”。语义特征提取过程为：对样本文本序列数据7012进行隐层特征提取(语义特征提取)，得到样本文本序列数据7012对应的样本文本语义信息7013；进一步地，基于样本全局语义符“[CLS]”对样本文本序列数据7012进行全局语义特征提取，得到该样本文本序列数据7012对应的样本全局隐状态“h [CLS]”；同时分别对样本文本序列数据7012中的N个样本对话词组进行上下文语义特征提取，得到N个样本对话词组分别对应的样本词组隐状态，也即得到N个样本词组隐状态，比如：样本对话词组“S 1”对应的样本词组隐状态
Figure PCTCN2021108503-appb-000003
(下标1表示样本文本序列数据7012中出现的第一个第一样本论元)，样本对话词组“Mom”对应的样本词组隐状态“h i”(i表示N个样本对话词组中的第i个样本对话词组)，样本对话词组“S 2”对应的样本词组隐状态
Figure PCTCN2021108503-appb-000004
同时对样本论元对进行语义特征提取，得到第一样本论元符号对应的第一初始样本论元隐状态
Figure PCTCN2021108503-appb-000005
第二样本论元符号对应的第二初始样本论元隐状态
Figure PCTCN2021108503-appb-000006
j表示样本文本序列数据7012中出现的第j个第一样本论元符号或第二样本论元符号；可选的，在样本文本序列数据7012中，第一样本论元符号与第二样本论元符号的数量可以相同，也可以不相同。其中，将样本全局隐状态、N个样本词组隐状态、第一初始样本论元隐状态及第二初始样本论元隐状态，确定为样本对话文本对应的样本文本语义信息。
步骤502，基于样本文本语义信息预测样本论元之间的预测论元关系。
在一种可能的实施方式中，当计算机设备获取到样本文本语义信息后，由于该样本文本语义信息中包含样本对话文本的全局语义信息，样本论元对应的上下文语义信息，对应的，可以基于样本文本语义信息来预测样本论元之间的论元关系，也即获取第一样本论元和第二样本论元之间的预测论元关系。
若需要预测样本论元对中第一样本论元和第二样本论元之间的论元关系，则首先需要获取到与样本论元相关的样本文本语义信息，由样本文本序列数据可知，样本论元出现在样本对话文本以及样本论元对中，对应的，就需要从样本文本语义信息中获取样本对话文本中各个样本论元对应的语义表示，以及样本论元对中各个样本论元对应的语义表示。
由于样本论元对包含第一样本论元和第二样本论元，对应的，确定与样本论元相关的样本文本语义信息的过程可以包括：确定第一样本论元对应的语义表示过程，以及确定第二样本论元对应的语义表示过程。在一个示例性的例子中，步骤502可以包含步骤502A～502F。
步骤502A，从样本文本语义信息所包括的N个样本词组隐状态中，获取第一样本论元符号对应的至少一个第一样本词组隐状态。
确定与第一样本论元相关的语义表示过程为：从样本文本语义信息中获取N个样本词组隐状态，并从N个样本词组隐状态中确定出第一样本论元符号(第一样本论元)对应的至少一个第一样本词组隐状态；同时，从样本文本语义信息中获取第一初始样本论元隐状态，进 而基于第一样本词组隐状态和第一初始样本论元隐状态确定第一样本论元对应的语义表示。
步骤502B，对各个第一样本词组隐状态与第一初始样本论元隐状态进行最大池化处理，得到第一样本论元符号对应的第一样本论元隐状态。
由于样本对话文本中包含多个第一样本论元，对应的，在进行语义特征提取过程中，也会对应提取到多个第一样本论元对应的第一样本词组隐状态，为了整合多个第一样本词组隐状态以及第一初始样本论元隐状态，在一种可能的实施方式中，对各个第一样本词组隐状态与第一初始样本论元隐状态进行最大池化处理，得到第一样本论元符号对应的第一样本论元隐状态。
换句话说，第一样本论元隐状态的确定过程为：从样本文本语义信息中获取第一样本论元符号对应的语义特征表示(包括样本对话文本部分提取到的第一样本词组隐状态，以及样本论元对部分提取到的第一初始样本论元隐状态)，通过对全部语义特征表示进行最大池化处理后，可以得到第一样本论元隐状态。
如图7所示，
Figure PCTCN2021108503-appb-000007
表示从样本文本语义信息中获取到的多个第一样本词组隐状态，
Figure PCTCN2021108503-appb-000008
表示第一初始样本论元隐状态，则对
Figure PCTCN2021108503-appb-000009
以及
Figure PCTCN2021108503-appb-000010
进行最大池化处理后，即可以得到第一样本论元符号对应的第一样本论元隐状态
Figure PCTCN2021108503-appb-000011
步骤502C，从N个样本词组隐状态中获取第二样本论元符号对应的至少一个第二样本词组隐状态。
确定与第二样本论元相关的语义表示过程为：从样本文本语义信息中获取N个样本词组隐状态，并从N个样本词组隐状态中确定出第二样本论元符号(第二样本论元)对应的至少一个第二样本词组隐状态；同时，从样本文本语义信息中获取第二初始样本论元隐状态，进而基于第二样本词组隐状态和第二初始样本论元隐状态，确定第二样本论元对应的语义表示。
步骤502D，对各个第二样本词组隐状态与第二初始样本论元隐状态进行最大池化处理，得到第二样本论元符号对应的第二样本论元隐状态。
由于样本对话文本中包含多个第二样本论元，对应的，在进行语义特征提取过程中，也会对应提取到多个第二样本论元对应的第二样本词组隐状态，为了整合多个第二样本词组隐状态以及第二初始样本论元隐状态，在一种可能的实施方式中，对各个第二样本词组隐状态与第二初始样本论元隐状态进行最大池化处理，得到第二样本论元符号对应的第二样本论元隐状态。
换句话说，第二样本论元隐状态的确定过程为：从样本文本语义信息中获取第二样本论元符号对应的语义特征表示(包括样本对话文本部分提取到的第二样本词组隐状态，以及样本论元对部分提取到的第二初始样本论元隐状态)，通过对全部语义特征表示进行最大池化处理后，可以得到第二样本论元隐状态。
如图7所示，
Figure PCTCN2021108503-appb-000012
表示从样本文本语义信息中获取到的多个第二样本词组隐状态，
Figure PCTCN2021108503-appb-000013
表示第二初始样本论元隐状态，则对
Figure PCTCN2021108503-appb-000014
以及
Figure PCTCN2021108503-appb-000015
进行最大池化处理后，即可以得到第二样本论元符号对应的第二样本论元隐状态
Figure PCTCN2021108503-appb-000016
步骤502E，对样本全局隐状态、第一样本论元隐状态及第二样本论元隐状态进行拼接，得到样本隐状态信息。
在进行样本论元关系预测过程中，不仅需要样本论元各自对应的语义特征表示(第一样本论元隐状态和第二样本论元隐状态)，还需要基于样本对话文本的全局语义进行关系预测，因此，在一种可能的实施方式中，从样本文本语义信息中获取到样本全局隐状态，进而基于样本全局隐状态、第一样本论元隐状态以及第二样本论元隐状态来共同进行论元关系预测。
可选的，为了将三个隐状态信息共同输入初始关系预测网络，需要对样本全局隐状态、第一样本论元隐状态，以及第二样本论元隐状态进行拼接，得到样本隐状态信息，进而将样本隐状态信息输出初始关系预测网络，进行论元关系预测。
在一个示例性的例子中，样本全局隐状态可以表示为h [CLS]，第一样本论元隐状态可以表 示为
Figure PCTCN2021108503-appb-000017
第二样本论元隐状态可以表示为
Figure PCTCN2021108503-appb-000018
则样本隐状态信息可以表示为
Figure PCTCN2021108503-appb-000019
Figure PCTCN2021108503-appb-000020
步骤502F，基于样本隐状态信息预测第一样本论元与第二样本论元之间的预测论元关系。
在一种可能的实施方式中，当获取到样本隐状态信息后，即可以将样本隐状态信息输入初始关系预测网络，由初始关系预测网络基于样本隐状态信息，预测第一样本论元和第二样本论元之间的预测论元关系。
如图7所示，在关系预测任务702中，基于样本隐状态信息7021(包括样本全局隐状态h [CLS]、第一样本论元隐状态
Figure PCTCN2021108503-appb-000021
以及第二样本论元隐状态
Figure PCTCN2021108503-appb-000022
)进行论元关系预测，得到第一样本论元和第二样本论元对应的预测论元关系。
在本申请实施例中，该初始关系预测模型包括初始关系预测网络，N个样本对话词组包括第一样本论元符号、第二样本论元符号，以及除第一样本论元和第二样本论元之外的其他样本对话词组。在基于样本文本语义信息预测第一样本论元与第二样本论元之间的预测论元关系时，计算机设备可以基于初始关系预测网络，从样本文本语义信息所包括的N个样本词组隐状态中，获取第一样本论元符号对应的第一样本词组隐状态，对第一样本词组隐状态与第一初始样本论元隐状态进行最大池化处理，得到第一样本论元符号的第一样本论元隐状态；从N个样本词组隐状态中获取第二样本论元符号对应的第二样本词组隐状态，对第二样本词组隐状态与第二初始样本论元隐状态进行最大池化处理，得到第二样本论元符号的第二样本论元隐状态。可选的，该第一样本论元隐状态及第二样本论元隐状态也可以是在共享编码器(或上文实施例中的初始语言表征网络)中生成的，在初始关系预测模型中，将共享编码器得到的第一样本论元隐状态及第二样本论元隐状态，输入初始关系预测网络。进一步地，计算机设备可以对样本全局隐状态、第一样本论元隐状态及第二样本论元隐状态进行拼接，得到样本隐状态信息；基于样本隐状态信息预测第一样本论元与第二样本论元之间的预测论元关系。
可选的，为了进一步提高论元关系的预测准确性，当计算机设备获取到样本隐状态信息后，还可以对样本隐状态信息进行语义增强，进而基于增强后的样本增强语义信息进行论元关系预测。
在一个示例性的例子中，步骤502F还可以包括以下步骤三～步骤五。
步骤三、对样本隐状态信息进行语义增强，得到样本增强语义信息。
其中，可以采用具备语义增强功能的网络对样本隐状态信息进行语义增强，比如，高速公路神经网络，或者其他具备语义增强功能的神经网络；换句话说，初始关系预测模型中还可以包括高速公路神经网络，该高速公路神经网络用于对提取到的语义信息进行语义增强。
在一种可能的实施方式中，将样本隐状态信息输入高速公路神经网络，由高速公路神经网络对其进行语义增强，以提取更深层次的语义特征，得到样本增强语义信息，进而基于该样本增强语义信息进行论元关系预测。
步骤四、基于样本增强语义信息，确定第一样本论元与第二样本论元对应的M种候选论元关系的样本关系预测概率，M为正整数。
其中，在初始关系预测模型训练之前，预先设置有M种候选论元关系，而初始关系预测网络的任务即是：基于输入的样本增强语义信息，预测第一样本论元和第二样本论元的预测论元关系，属于M种候选论元关系的样本关系预测概率。
在一种可能的实施方式中，将样本增强语义信息输入初始关系预测网络中，由初始关系预测网络基于样本增强语义信息中的全局语义信息、上下文语义信息，以及论元相关语义信息，预测第一样本论元和第二样本论元属于各种候选论元关系的样本关系预测概率。
步骤五、将最大样本关系预测概率对应的候选论元关系，确定为第一样本论元与第二样本论元之间的预测论元关系。
当获取到初始关系预测网络输出的各个样本关系预测概率后，默认将最大样本关系预测 概率对应的候选论元关系，确定为第一样本论元和第二样本论元之间的预测论元关系。
进一步地，在基于样本隐状态信息预测第一样本论元与第二样本论元之间的预测论元关系时，对样本隐状态信息进行语义增强，得到样本增强语义信息。可选的，计算机设备可以基于融合网络对样本隐状态信息进行语义增强，该融合网络可以是任意一个可以进行语义增强的网络，如包括一个或至少两个高速公路网络(highway network)的网络等，基于样本增强语义信息，预测第一样本论元与第二样本论元对应的M种候选论元关系的样本关系预测概率，将具有最大的样本关系预测概率的候选论元关系，确定为第一样本论元与第二样本论元之间的预测论元关系，M为正整数。
可选的，在其他可能的实施方式中，可以无需对样本隐状态信息进行语义增强，即将样本隐状态信息直接输入初始关系预测网络中，由初始关系预测网络基于样本隐状态信息中包含的全局语义信息、上下文语义信息，以及论元相关的语义信息等，预测第一样本论元和第二样本论元之间的预测论元关系。
步骤503，根据样本论元之间的实际论元关系与预测论元关系生成第一损失。
为了使得初始关系预测模型可以学习到更有效的语义特征，需要基于每轮的预测损失对模型参数进行更新，而对于论元关系预测任务，在一种可能的实施方式中，可以根据样本论元对对应的实际论元关系和预测论元关系，生成论元关系预测损失，也即第一损失，以便后续可以基于第一损失更新初始关系预测模型的模型参数。
进一步地，计算机设备可以根据第一样本论元与第二样本论元之间的实际论元关系与预测论元关系生成第一损失，该第一损失可以是实际论元关系与预测论元关系之间的二值交叉熵损失、对数损失或平方损失等，在此不做限制。
具体的，图7所示，从样本文本语义信息所包括的N个样本词组隐状态中，获取第一样本论元符号S 1对应的第一样本词组隐状态，即
Figure PCTCN2021108503-appb-000023
等，对第一样本词组隐状态与第一初始样本论元隐状态
Figure PCTCN2021108503-appb-000024
进行最大池化处理，得到第一样本论元符号的第一样本论元隐状态
Figure PCTCN2021108503-appb-000025
从N个样本词组隐状态中获取第二样本论元符号S 2对应的第二样本词组隐状态，即
Figure PCTCN2021108503-appb-000026
等，对第二样本词组隐状态与第二初始样本论元隐状态
Figure PCTCN2021108503-appb-000027
进行最大池化处理，得到第二样本论元符号的第二样本论元隐状态
Figure PCTCN2021108503-appb-000028
在关系预测任务702中，对样本全局隐状态h [CLS]、第一样本论元隐状态
Figure PCTCN2021108503-appb-000029
及第二样本论元隐状态
Figure PCTCN2021108503-appb-000030
进行拼接，得到样本隐状态信息7021，基于样本隐状态信息7021预测第一样本论元与第二样本论元之间的预测论元关系。其中，该样本隐状态信息7021可以记作h，
Figure PCTCN2021108503-appb-000031
该样本全局隐状态表示了该样本文本序列数据中的全局语义信息，融合了样本文本序列数据中的各个词组分别与样本全局语义符之间的关系，第一样本论元隐状态表示与第一样本论元相关的局部语义信息，第二样本论元隐状态表示与第二样本论元相关的局部语义信息，将样本全局隐状态h [CLS]、第一样本论元隐状态
Figure PCTCN2021108503-appb-000032
及第二样本论元隐状态
Figure PCTCN2021108503-appb-000033
进行拼接，得到样本隐状态信息7021，使得样本隐状态信息可以包括样本文本序列数据中的全局语义信息及局部语义信息等，可以更为全面地提取该样本文本序列数据中的特征，从而提高关系预测的准确性。
步骤504，获取样本对话文本中的隐藏字符。
由于训练样本(样本对话文本以及样本论元对)不仅需要用于论元关系预测任务，还用于说话者预测任务(隐藏字符预测任务)，而说话者预测任务的任务目的是为了预测样本对话文本中被遮盖或隐藏的字符，因此，在一种可能的实施方式中，输入初始关系预测模型的样本对话文本是经过遮盖或隐藏处理之后的对话文本，也就是说，存在原始对话文本，对原始对话文本进行隐藏处理之后，可以得到样本对话文本。
其中，对原始对话文本进行隐藏处理，得到样本对话文本的过程可以包括步骤六～步骤八。
步骤六、获取原始对话文本和样本论元对。
其中，原始对话文本：未经过隐藏处理的对话文本。样本论元对中的各个样本论元也属于原始对话文本。
步骤七、响应于样本论元对中存在至少一个样本论元的论元类型为说话者论元，基于样本论元对确定隐藏字符。
由于说话者预测任务(初始词组预测网络)是用于预测被掩盖的说话者，以便使得初始关系预测模型学习到说话者特征，进而辅助与说话者相关的论元关系预测，对应需要基于样本论元对中的说话者论元，确定掩盖原始对话文本中的对应字符，因此，在一种可能的实施方式中，当样本论元对中存在至少一个说话者论元，可以基于该说话者论元确定原始样本对话文本中需要隐藏的隐藏字符。
可选的，若样本论元对中包含的两个样本论元的论元类型均为说话者论元，则可以随机选取任意样本论元，并将样本论元在原始对话文本中对应的词组确定为隐藏字符；可选的，若样本论元对中仅包含单个说话者论元，则直接将该说话者论元在原始对话文本中对应的词组确定为隐藏字符。
步骤八、基于隐藏字符对原始对话文本进行隐藏处理，得到样本对话文本。
在一种可能的实施方式中，当确定出隐藏字符后，则可以对原始对话文本中该隐藏字符所对应的词组进行隐藏处理，进而将隐藏处理后的原始对话文本，确定为样本对话文本，进行后续语义特征提取、论元关系预测、触发词预测、隐藏字符预测等任务。
在一个示例性的例子中，原始对话文本为：d＝s 1：t 1，s 2：t 2，s 1：t i…，s m：t m，样本论元对为(s 1，s 2)，该样本论元对中包含的各个样本论元均为说话者论元，可以将s 1确定为隐藏字符，对应的，经过隐藏处理后得到的样本对话文本可以表示为：d＝s 1：t 1，s 2：t 2，C：t i…，s m：t m，其中，C表示将样本论元s 1隐藏处理后的符号。
可选的，隐藏处理可以是将该隐藏字符用其他无意义字符替换，也可以是对隐藏字符进行乱码处理，或采用其他隐藏方式，本申请实施例对此不构成限定。
需要说明的是，由于原始对话文本中样本论元还需要参与后续论元关系预测过程，因此，在基于隐藏字符对原始对话文本进行隐藏处理时，不能对原始对话文本中包含的全部隐藏字符(样本论元)进行隐藏处理，而是以预设概率随机掩盖原始对话文本中的隐藏字符。
示意性的，预设概率可以为10％，也即对原始对话文本中隐藏字符的10％进行随机掩盖，比如，原始对话文本中包含10个隐藏字符，则从10个隐藏字符中随机选择1个隐藏字符，进行隐藏处理，得到样本对话文本。
初始词组预测网络需要基于样本文本语义信息，预测样本对话文本中被遮盖或隐藏的隐藏字符，对应的，需要预先输入隐藏字符(实际隐藏字符)，使得计算机设备可以获取到样本对话文本中的隐藏字符，以便后续基于隐藏字符预测结果和隐藏字符，计算隐藏字符预测损失。
步骤505，基于样本文本语义信息，预测隐藏字符对应的预测字符。
本申请实施例中，对于初始语言表征网络输出的样本文本语义信息，分别由初始关系预测网络、初始词组预测网络以及初始触发词预测网络共享，因此，在一种可能的实施方式中，初始词组预测网络也可以基于样本文本语义信息，预测样本对话文本中被隐藏的隐藏字符，得到预测字符。
在隐藏字符预测过程中，仅需要样本对话文本中隐藏字符所对应的语义特征表示，在一个示例性的例子中，步骤505可以包括步骤505A和步骤505B。
步骤505A，从样本文本语义信息中确定隐藏字符对应的掩码隐状态，掩码隐状态用于表征隐藏字符在样本对话文本中对应的语义信息。
由于在隐藏字符预测任务(说话者预测任务)中，需要基于提取到的语义信息来预测隐藏字符，而预测隐藏字符仅需要隐藏字符相关的语义信息即可，因此，在一种可能的实施方式中，需要从样本文本语义信息中获取隐藏字符所对应的样本词组隐状态，并将其确定为掩码隐状态，以便使用该掩码隐状态进行隐藏字符预测。
步骤505B，基于掩码隐状态，预测隐藏字符对应的预测字符。
由于掩码隐状态可以表征隐藏字符在样本对话文本中对应的语义信息，因此，基于该掩码隐状态，即可预测出预测字符。
在一个示例性的例子中，若样本对话文本可以表示为：d＝s 1：t 1，s 2：t 2，C：t i…，s m：t m，其中，C表示被隐藏的说话者论元，对样本对话文本和样本论元对进行数据预处理，并输入初始语言表征网络后，可以得到初始语言表征网络输出的样本文本语义信息，该样本文本语义信息中包含有“C”这个隐藏字符所对应的掩码隐状态
Figure PCTCN2021108503-appb-000034
进而将该掩码隐状态
Figure PCTCN2021108503-appb-000035
输入初始词组预测网络，进行隐藏字符预测，得到预测字符。如图7所示，在词组预测任务703中，从样本文本语义信息7013中获取到隐藏字符对应的掩码隐状态
Figure PCTCN2021108503-appb-000036
将该掩码隐状态
Figure PCTCN2021108503-appb-000037
输入初始词组预测网络，可以得到掩码隐状态
Figure PCTCN2021108503-appb-000038
对应的预测字符。
步骤506，根据隐藏字符与预测字符生成第二损失。
为了使得初始关系预测模型更有效学习到说话者相关特征，以进一步辅助说话者论元的关系预测，因此，在一种可能的实施方式中，基于隐藏字符和初始词组预测网络输出的预测字符，确定隐藏字符预测损失，也即第二损失，用于更新初始关系预测模型。
进一步地，计算机设备可以根据隐藏字符与预测字符生成第二损失，该第二损失可以是预测字符与隐藏字符之间的二值交叉熵损失、对数损失或平方损失等，在此不做限制。
步骤507，根据实际论元关系与样本文本语义信息生成触发词检测文本数据。
在本申请实施例中，该初始关系预测模型还包括初始触发词预测网络，其中，触发词是指文本中能够清晰指示论元关系的词语，触发词预测网络用于根据样本文本语义信息，预测样本对话文本中的触发词。
初始触发词预测网络在判断样本对话文本中的各个样本对话词组是否为触发词时，即判断各个样本对话词组是否与论元关系存在特定联系时，需要获取到实际论元关系，以及各个样本对话词组的语义表示，因此，在一种可能的实施方式中，可以根据实际论元关系和样本文本语义信息，生成触发词检测文本数据，并将该触发词检测文本数据输入初始触发词预测网络中，由初始触发词预测网络预测样本对话文本中的触发词。
在一个示例性的例子中，步骤507可以包括步骤507A和步骤507B。
步骤507A，确定实际论元关系对应的论元关系向量。
由于样本文本语义信息为特征向量的形式，为了可以将样本文本语义信息与实际论元关系进行拼接，在一种可能的实施方式中，将实际论元关系也转化为特征向量的形式，即得到实际论元关系对应的论元关系向量。
步骤507B，对论元关系向量和样本文本语义信息进行拼接，生成触发词检测文本数据。
当得到论元关系向量后，即可以将论元关系向量和样本文本语义信息进行拼接，生成触发词检测文本数据，用于后续的触发词预测过程。
在触发词预测任务中，初始触发词预测网络需要识别样本对话文本中的每个样本对话词组是否为触发词，对应在进行拼接时，需要将论元关系向量与样本文本语义信息中的各个样本词组隐状态进行拼接；在一个示例性的例子中，步骤507B还可以包括以下步骤九～步骤十一。
步骤九、从样本文本语义信息中确定出至少一个样本词组隐状态，样本词组隐状态用于表征样本对话词组在样本对话文本中对应的语义信息。
其中，样本对话词组为样本对话文本中包含的各个字符。示例性的，若样本对话文本为：a 1:Mom！a 2：Sweetie……”，其中，样本对话词组包括：a 1、Mom、a 2、Sweetie等。
为了使得初始触发词预测网络可以识别样本对话文本中的各个样本对话词组是否为触发词，在一种可能的实施方式中，从样本文本语义信息中确定出各个样本对话词组对应的语义信息，也即样本词组隐状态，进而将该样本词组隐状态与论元关系向量进行拼接，用于后续预测该样本对话词组是否为触发词。
步骤十、将论元关系向量与样本词组隐状态进行拼接，得到样本对话词组对应的触发词 检测文本。
针对每个样本词组隐状态，将该样本词组隐状态与论元关系向量进行拼接，得到该样本对话词组对应的触发词检测文本，也就是说触发词检测文本中包含论元关系向量和一个样本词组隐状态。
如图7所示，从样本文本语义信息7013中获取到样本对话词组“Mom”对应的样本词组隐状态“h i”，将该样本词组隐状态“h i”与论元关系向量“e r”进行拼接，得到样本对话词组“Mom”对应的触发词检测文本“[e r，h i]”。
步骤十一、将各个样本对话词组对应的触发词检测文本，确定为触发词检测文本数据。
为了提高初始触发词的预测效率，在一种可能的实施方式中，将各个样本对话词组对应的样本词组隐状态，均与论元关系向量进行拼接，得到各个样本对话词组对应的触发词检测文本，进而将触发词检测文本的集合确定为触发词检测文本数据，共同输入初始触发词预测网络，由初始触发词预测网络同时预测各个样本对话词组是否为触发词。
在一个示例性的例子中，若样本对话文本中包含5个样本对话词组，获取各个样本对话词组对应的样本词组隐状态：h 1、h 2、h 3、h 4、h 5，将各个样本词组隐状态与论元关系向量进行拼接，得到5个触发词检测文本：[e r，h 1]、[e r，h 2]、[e r，h 3]、[e r，h 4]、[e r，h 5]，5个触发词检测文本即构成触发词检测文本数据，用于输入初始触发词检测网络进行触发词预测。
步骤508，预测触发词检测文本数据对应的预测序列标注。
在一种可能的实施方式中，计算机设备将触发词检测文本数据输入初始触发词预测网络中，由初始触发词预测网络基于各个样本对话词组的语义信息，以及实际论元关系，判断该样本对话词组是否与实际论元关系存在特定联系，从而确定出该样本对话词组是否为触发词。
在一个示例性的例子中，步骤508还可以包括步骤508A和步骤508B。
步骤508A，基于触发词检测文本数据进行触发词预测，得到各个样本对话词组对应的预测词组标注，预测词组标注用于表征样本对话词组所属的触发词类型。
本申请实施例中，在初始触发词预测网络对样本对话词组进行触发词预测的过程中，采用BIO标注模式标注样本对话文本，也就是说，由触发词预测网络输出的是由BIO构成的预测序列标注，其中，I表示该位置处的样本对话词组为触发词的中间字符，B表示该位置处的样本对话词组为触发词的起始字符，O表示该位置处的样本对话词组不是触发词。
可选的，还可以采用其他标注模式，比如：BIOES标注模式，其中，B表示该位置处的样本对话词组为触发词的起始字符，I表示该位置处的样本对话词组为触发词的中间字符，O表示该位置处的样本对话词组不是触发词，E表示该位置处的样本对话词组为触发词的结束字符，S表示该样本对话词组为单个字符(与触发词无关)。
在一种可能的实施方式中，将触发词检测文本数据输入初始触发词预测网络，进行触发词预测，可以得到各个样本对话词组对应的预测词组标注，也即确定该样本对话词组是否是触发词，以及是否是触发词的起始字符、中间字符或结束字符。
可选的，初始词组预测网络在进行触发词预测过程中，会预测该样本对话词组对应多种候选词组标注的概率，其中，若标注模式采用BIO标注法，则候选词组标注为三种：B、I、O；若标注模式采用BIOES标注法，则候选词组标注为五种：B、I、O、E、S；进而将最大概率对应的候选词组标注确定为该样本对话词组对应的预测词组标注。
步骤508B，将各个样本对话词组对应的预测词组标注，确定为预测序列标注。
各个样本对话词组对应的预测词组标注，即可以构成样本对话文本对应的预测序列标注，也即预测序列标注为各个预测词组标注的集合。
如图7所示，触发词检测文本数据为：[e r，h 1]、[e r，h 2]…[e r，h i]、[e r，h i+1]、[e r，h i+2]，将触发词检测文本数据输入初始触发词预测网络，得到的预测标注序列为：O、O…B、I、O，可见，第一个样本对话词组、第二个样本对话词组和第i+2个样本对话词组对应的预 测标注均为O，表示第一个对话样本词组、第二个对话样本词组和第i+2个样本对话词组均不是触发词，而第i样本对话词组对应的预测标注为B，表示第i样本对话词组为触发词的起始字符，第i+1样本对话词组对应的预测标注为I，表示第i+1样本对话词组为触发词的中间字符，且第i+2样本对话词组不是触发词，则基于该预测标注序列可知，样本对话文本中的预测触发词由第i样本对话词组和第i+1样本对话词组构成。
步骤509，根据触发词检测文本数据对应的实际序列标注与预测序列标注生成第三损失。
为了使得初始关系预测模型更有效的学习与触发词相关的语义特征，在一种可能的实施方式中，基于触发词检测文本数据对应的实际序列标注和预测序列标注，确定触发词预测损失，即第三损失，用于更新初始关系预测模型。
可选的，实际序列标注由人工预先标注，并将实际序列标注与对应的样本对话文本关联存储在计算机设备中，以便计算机设备在计算第三损失时，可以获取到对应的实际序列标注。
进一步地，计算机设备可以根据触发词检测文本数据中的实际序列标注与预测序列标注生成第三损失，该第三损失可以是实际序列标注与预测序列标注之间的二值交叉熵、对数损失或平方损失等，在此不做限制。
需要说明的是，在初始关系预测模型的模型训练过程中，使用样本论元对对应的实际论元关系生成触发词检测文本数据；而在模型测试阶段，为了验证初始关系预测模型在三个任务中的预测准确性，采用初始关系预测网络输出的预测论元关系，生成触发词检测文本数据，也就是说，计算机设备可以获取初始关系预测网络中得到的预测论元关系，对预测论元关系进行向量转换，得到预测论元关系对应的预测关系向量，将预测关系向量与N个样本词组隐状态进行拼接，生成触发词测试数据，预测触发词测试数据中的测试序列标注，基于测试序列标注与实际序列标注生成的第四损失，对初始关系预测网络进行优化调整。
步骤510，根据第一损失、第二损失及第三损失对初始关系预测模型进行模型参数调整，生成对话关系预测模型。
在一种可能的实施方式中，根据第一损失、第二损失和第三损失确定联合损失，进而基于该联合损失训练初始关系预测模型，更新初始关系预测模型中各个子网络的模型参数，经过多轮训练后，直至损失收敛，得到对话关系预测模型。
可选的，在训练结束后，可以使用测试样本集对对话关系预测模型进行模型测试，从而进一步优化对话关系预测模型。
需要说明的是，在初始对话关系预测模型的训练过程中，需要由初始词组预测网络和初始触发词预测网络进行辅助训练，而当初始对话关系预测模型训练完成后，为了简化模型，可以删除初始对话关系预测模型中的初始词组预测网络和初始触发词预测网络，得到对话关系预测模型，也即对话关系预测模型中可以仅包含语言表征网络和关系预测网络；可选的，对话关系预测模型中还可以包含语义增强网络。
在本申请实施例中，对话关系预测模型用于预测目标对话文本中的目标论元对之间的目标论元关系。具体的，计算机设备可以根据第一损失、第二损失及第三损失生成模型损失，基于模型损失对初始关系预测模型进行模型参数调整，具体是对初始关系预测模型中的各个网络进行模型参数调整，生成对话关系预测模型。可选的，将第一损失记作L 1，将第二损失记作L 2，将第三损失记作L 3，基于第一损失、第二损失及第三损失得到模型损失，该模型损失为“L＝L 1+L 2+L 3”，基于模型损失对初始关系预测模型进行模型参数调整。通过上述过程，对关系预测任务、词组预测任务及触发词预测任务等以多任务学习的方式进行联合训练，使得该词组预测任务及触发词预测任务等的信息可以提供至关系预测任务，即该关系预测任务可以基于词组预测任务得到论元的特征信息等，可以基于触发词预测任务得到触发词信息，从而使得该关系预测任务可以更为全面地获取到样本对话文本中的信息，提高了关系预测的准确性。其中，多任务学习(Multi-task Learning，MTL)，是一种同时考虑多个相关任务的机器学习方法，利用各个任务之间的内在关系提高单个任务学习的泛化性能，即，利用关系预 测任务、词组预测任务及触发词预测任务等之间的内在关系，提高关系预测任务的泛化性能。
在本申请实施例中，将样本对话文本及样本对话文本中的样本论元对输入初始关系预测模型，在初始关系预测模型中对样本对话文本及样本论元对进行特征提取，得到样本对话文本对应的样本文本语义信息；样本论元对包括第一样本论元及第二样本论元；基于样本文本语义信息预测第一样本论元与第二样本论元之间的预测论元关系，根据第一样本论元与第二样本论元之间的实际论元关系与预测论元关系生成第一损失函数；获取样本对话文本与样本论元对中的隐藏字符，预测隐藏字符对应的预测字符，根据隐藏字符与预测字符生成第二损失函数；根据实际论元关系与样本文本语义信息生成触发词检测文本数据，预测触发词检测文本数据中的预测序列标注，根据触发词检测文本数据中的实际序列标注与预测序列标注生成第三损失函数；根据第一损失函数、第二损失函数及第三损失函数对初始关系预测模型进行模型参数调整，生成对话关系预测模型；对话关系预测模型用于预测目标对话文本中的目标论元对之间的目标论元关系。通过上述过程，本申请将对话式关系抽取分为三个相关的子任务，分别是关系预测任务、词组预测任务及触发词预测任务，通过综合这三个子任务，对模型进行联合训练，可以充分利用从词组预测任务及触发词预测任务中学到的有效信息，并基于该有效信息影响关系预测任务，从而提高对话关系处理的准确性。
上文实施例主要描述了初始关系预测模型的训练过程，当初始关系预测模型训练完成，生成对话关系预测模型后，该对话关系预测模型即可用于不同对话关系预测场景。
进一步地，可以参见图8，图8是本申请一个示例性实施例提供的对话关系处理方法的流程图。本申请实施例以该方法应用于图1所示的计算机设备为例进行说明，该方法包括如下步骤：
步骤801，将目标对话文本及目标论元对输入对话关系预测模型，在对话关系预测模型中对目标对话文本及目标论元对进行语义特征提取，得到目标对话文本对应的目标文本语义信息，目标论元对中的各个目标论元属于目标对话文本。
其中，对话关系预测模型中可以仅包含语言表征网络和关系预测网络；其中，语言表征网络用于提取目标对话文本对应的目标文本语义信息，而关系预测网络用于预测目标论元对中各个目标论元之间的目标论元关系；可选的，对话关系预测模型还可以包含语义增强网络，用于对目标文本语义信息进行语义增强。
在一种可能的实施方式中，将目标对话文本和目标论元对输入对话关系预测模型中的语言表征网络，由语言表征网络对其进行语义特征提取，以得到目标对话文本对应的目标文本语义信息。
其中，在进行语义特征提取之前，还需要对目标对话文本和目标论元对进行数据预处理，该过程可以参考上文实施例中样本对话文本和样本论元对的数据预处理过程，本申请实施例在此不做赘述。
在本申请实施例中，计算机设备基于目标文本拼接符对目标对话文本与目标论元对进行拼接，生成目标拼接文本数据；将目标拼接文本数据中的第一目标论元替换为第一目标论元符号，将目标拼接文本数据中的第二目标论元替换为第二目标论元符号，生成目标文本序列数据；对目标文本序列数据进行语义特征提取，得到目标对话文本对应的目标文本语义信息。其中，该目标文本拼接符与样本文本拼接符相同，只是处于不同的阶段时的名称不同，例如，样本文本拼接符包括样本全局语义符及样本分隔符等，假定样本全局语义符为[CLS]，样本分隔符为[SEP]，则该目标文本拼接符包括目标全局语义符及目标分隔符等，该目标全局语义符为[CLS]，目标标分隔符为[SEP]。进一步地，该目标文本序列数据包括目标对话文本对应的目标对话序列数据，目标对话序列数据包括v个目标对话词组，v为正整数，计算机设备可以对目标文本序列数据中的目标全局语义符、v个目标对话词组、第一目标论元符号及第二目标论元符号分别进行隐层特征提取，得到目标全局语义符对应的目标全局隐状态、N个目标对话词组分别对应的目标词组隐状态、第一目标论元符号对应的第一初始目标论元隐状态 以及第二目标论元符号对应的第二初始目标论元隐状态；将目标全局隐状态、v个目标词组隐状态、第一初始目标论元隐状态及第二初始目标论元隐状态，确定为目标对话文本对应的目标文本语义信息。
在模型应用阶段，目标对话文本和目标论元对可以是由用户自行输入的，也即用户将目标对话文本，以及需要提取论元关系的至少一个目标论元对输入计算机设备，由计算机设备获取到该目标对话文本和目标论元对；在其他可能的实施方式中，目标论元对可能需要计算机设备从对话咨询信息中获取得到，该对话咨询信息可以是：用户咨询语句、阅读理解问题等。
可选的，计算机设备可以直接获取目标用户提供的目标对话文本及目标对话文本中的目标论元对。或者，计算机设备可以获取目标对话文本与目标对话文本关联的对话咨询信息，对对话咨询信息进行解析，提取对话咨询信息中的目标论元对；例如，获取到对话咨询信息为“What is the relationship between Speaker 2and Speaker 4？”，对对话咨询信息进行解析，提取该对话咨询信息中的目标论元对，可以得到，目标论元对包括第一目标论元“Speaker 2”及第二目标论元“Speaker 4”。可选的，若计算机设备在对话咨询信息中获取到第一目标论元，则解析对话咨询信息，获取关联论元类型，从目标对话文本中获取该关联论元类型对应的论元，将该关联论元类型对应的论元确定为第二目标论元，其中，该关联论元类型包括但不限于人物类型、物品类型或动物类型等，该第二目标论元的数量为一个或至少两个。基于对话关系预测模型，获取第一目标论元与每个第二目标论元之间的目标论元关系。
步骤802，基于对话关系预测模型中的关系预测网络对目标文本语义信息进行预测，得到第一目标论元与第二目标论元之间的目标论元关系。
在一种可能的实施方式中，当计算机设备获取到目标对话文本对应的目标文本语义信息后，可以从目标文本语义信息中确定出目标全局语义信息、第一目标论元对应的第一目标论元隐状态，以及第二目标论元对应的第二目标论元隐状态，并将目标全局语义信息、第一目标论元隐状态，以及第二目标论元隐状态，确定为目标隐状态信息，输入关于预测网络，由关系预测网络进行论元关系预测，得到第一目标论元和第二目标论元之间的目标论元关系。
其中，第一目标论元隐状态的确定过程，可以参考上文实施例中第一样本论元隐状态的确定过程，第二目标论元隐状态的确定过程，也可以参考上文实施例中第二样本论元隐状态的确定过程，本申请实施例此在不做赘述。
在本申请实施例中，对话关系预测模型是基于第一损失、第二损失及第三损失对初始关系预测模型训练得到的；初始关系预测模型包括初始关系预测网络、初始词组预测网络及初始触发词预测网络；第一损失函数是由第一样本论元与第二样本论元之间的实际论元关系与预测论元关系生成的，实际论元关系是在初始关系预测网络中，对第一样本论元与第二样本论元所在的样本对话文本对应的样本文本语义信息预测得到的；第二损失是由隐藏字符与预测字符生成的，预测字符是在初始词组预测网络中，对样本对话文本中的隐藏字符进行预测得到的；第三损失是由触发词检测文本数据中的实际序列标注与预测序列标注生成的，触发词检测文本数据是在初始触发词预测网络中，根据实际论元关系与样本文本语义信息生成的，预测序列标注是对触发词检测文本数据进行预测得到的。
其中，计算机设备可以对初始关系预测模型进行模型参数调整后，删除词组预测任务区域及触发词预测任务区域，得到对话关系预测模型，可以精简模型，减少对话关系预测模型所占用的资源。或者，计算机设备对初始关系预测模型进行模型参数调整后，保留词组预测任务区域及触发词预测任务区域，得到对话关系预测模型，便于后续对对话关系预测模型进行进一步地优化，此时，该对话关系预测模型的模型结构可以如图7所示。其中，可以参见图6，计算机设备在训练阶段，基于关系预测任务、词组预测任务及触发词预测任务进行联合训练，生成对话关系预测模型；在测试阶段，直接基于关系预测任务得到关系预测结果，该关系预测结果用于表示输入对话关系预测模型的目标论元对之间的论元关系。
进一步地，计算机设备可以根据目标对话文本、对话咨询信息及目标论元对生成目标问答数据，将目标问答数据添加至问答数据库中。
可选的，若计算机设备在对话咨询信息中获取到第一目标论元，则解析对话咨询信息，获取关联论元类型。具体的，若计算机设备在对话咨询信息中获取到目标论元关系，则获取该目标论元关系对应的关联论元类型，从目标对话文本中获取关联论元类型对应的候选论元，基于对话关系预测模型预测第一目标论元与候选论元之间的候选论元关系，将候选论元关系为目标论元关系的候选论元确定为第二目标论元，该第一目标论元与第二目标论元组成目标论元对，其中，该第二目标论元为对话咨询信息的回复数据，该候选论元的数量为一个或至少两个。例如，该对话咨询信息为“Who is the boss of Speaker 2？(谁是用户2的老板)”，在对话咨询信息中获取到第一目标论元“Speaker 2”，获取到目标论元关系为“boss”，该目标论元关系为“boss(老板)”对应的关联论元类型为人物类型，假定计算机设备从目标对话文本中获取人物类型对应的候选论元，假定该候选论元包括“Speaker 1、Speaker 3及Speaker 4”，基于对话关系预测模型，预测第一目标论元分别与各个候选论元之间的候选论元关系，假定第一目标论元与候选论元“Speaker 1”之间的候选论元关系为“subordinate(下属)”，第一目标论元与候选论元“Speaker 3”之间的候选论元关系为“friend(朋友)”，第一目标论元与候选论元“Speaker 4”之间的候选论元关系为“boss”，则将候选论元“Speaker 4”确定为第二目标论元。该第二目标论元为对话咨询信息的回复数据，即Speaker 4是Speaker 2的老板。
具体的，本申请可以应用于文旅领域中的信息抽取业务，进行关系预测能力的优化，可以快速地从介绍文档中提取出有效的知识信息，定制私有化知识图谱；或者，可以应用于问答系统，帮助完善问答系统的知识库，保障问答系统可以更有效地回答用户提问等，在此不做限制。
进一步地，在模型训练阶段，该初始关系预测模型的输入参数还可以包括任务类型、任务版本、使用区域、输入数据路径、输出数据路径及任务标识等，在此不做限制，具体可以参见表1所示：
表1
Figure PCTCN2021108503-appb-000039
其中，在表1中，任务类型(Action)用于表示本次模型预测任务的任务类型是训练类型还是测试类型等，此时的任务类型可以是训练类型；任务版本(Version)用于表示所使用的模型的版本号或版本生成时间等；使用区域(Region)用于表示可以应用该模型的地域列表、产品类型或用户列表等；输入数据路径(InputDatasetPath)用于表示训练模型等所使用的训练样本等的存储位置路径；输出数据路径(OutputModelPath)用于表示待训练模型的存储位置路径等；任务标识(ProjectId)用于表示本次模型预测任务的标识(ID)。可选的，该 表1中的各个输入参数可以不输入初始关系预测模型中，只作为该次模型训练阶段的一个日志记录。
其中，该初始关系预测模型的输出参数还可以包括训练完成时间及任务请求标识等，在此不做限制，具体可以参见表2所示：
表2
Figure PCTCN2021108503-appb-000040
其中，在表2中，训练完成时间(TimeOfTrain)用于表示本次模型预测任务完成所用的时长或完成时的时间；任务请求标识(RequestId)用于表示请求本次模型预测任务的请求标识(ID)，是请求的唯一标识符，可以用于表示任务的场景变化、发起请求的用户标识或请求的次数等。
进一步地，在模型测试或应用阶段，该对话关系预测模型的输入参数还可以包括任务类型、任务版本、使用区域、输入数据路径、输出数据路径及任务标识等，在此不做限制，具体可以参见表3所示：
表3
Figure PCTCN2021108503-appb-000041
其中，在表3中，任务类型(Action)用于表示本次模型预测任务的任务类型是训练类型还是测试类型等，此时的任务类型可以为测试类型或预测类型；Version用于表示所使用的模型的版本号或版本生成时间等；Region用于表示可以应用该模型的地域列表、产品类型或用户列表等；输入数据路径(InputDatasetPath)用于表示用于测试模型或使用模型等所使用的数据的存储位置路径，如目标对话文本等的存储路径；输出数据路径(OutputModelPath)用于表示已训练好的模型的存储位置路径等；任务标识(ProjectId)用于表示本次模型预测任务的标识(ID)。可选的，该表3中的各个输入参数可以不输入对话关系预测模型中，只作为该次模型训练阶段的一个日志记录。
其中，该对话关系预测模型的输出参数还可以包括输出结果数据路径及任务请求标识等，在此不做限制，具体可以参见表4所示：
表4
Figure PCTCN2021108503-appb-000042
其中，在表4中，输出结果数据路径(OutputDatasetPath)用于表示本次模型预测任务所得到的数据的存储位置路径，如目标论元关系等，该路径可以是区块链网络中的位置路径，或者是计算机设备中的存储路径，还可以是云存储技术所指示的存储路径等，在此不做限制；任务请求标识(RequestId)用于表示请求本次模型预测任务的请求标识(ID)，是请求的唯一标识符，可以用于表示任务的场景变化、发起请求的用户标识或请求的次数等。
请参见图9，图9是本申请一个示例性实施例提供的对话关系处理装置的结构示意图。该对话关系处理装置可以是运行于计算机设备中的一个计算机程序(包括程序代码等)，例如该对话关系处理装置可以为一个应用软件；该装置可以用于执行本申请实施例提供的方法中的相应步骤。如图9所示，该对话关系处理装置900可以包括：模型输入模块11、语义信息提取模块12、关系处理模块13、第一损失生成模块14、字符获取模块15、词组预测模块16、第二损失生成模块17、触发检测生成模块18、序列标注模块19、第三损失生成模块20及模型训练模块21。
模型输入模块11和语义信息提取模块12，用于通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息，所述样本论元对中的各个样本论元属于所述样本对话文本；
关系处理模块13和第一损失生成模块14，用于基于所述样本文本语义信息进行论元关系预测，并基于论元关系预测结果确定第一损失，所述论元关系预测结果用于表征所述样本论元之间的关系；
字符获取模块15、词组预测模块16以及第二损失生成模块17，用于基于所述样本文本语义信息进行隐藏字符预测，并基于隐藏字符预测结果确定第二损失，所述隐藏字符预测结果用于表征所述样本对话文本中的隐藏字符；
触发检测生成模块18、序列标注模块19以及第三损失生成模块20，用于基于所述样本文本语义信息和实际论元关系进行触发词预测，并基于触发词预测结果确定第三损失，所述实际论元关系为所述样本论元对应的标注论元关系，所述触发词预测结果用于表征触发词在所述样本对话文本中的位置；
模型训练模块21，用于基于所述第一损失、所述第二损失和所述第三损失，训练所述初始关系预测模型，得到对话关系预测模型。
在一种可选的实施方式中，模型输入模块11和语义信息提取模块12，还用于将所述样本对话文本及所述样本论元对输入所述初始关系预测模型，在所述初始关系预测模型中对所述样本对话文本及所述样本论元对进行语义特征提取，得到所述样本文本语义信息；
关系处理模块13，用于基于所述样本文本语义信息预测所述样本论元之间的预测论元关系；
第一损失生成模块14，用于根据所述样本论元之间的所述实际论元关系与所述预测论元关系生成所述第一损失；
字符获取模块15，用于获取所述样本对话文本中的隐藏字符；
词组预测模块16，用于基于所述样本文本语义信息，预测所述隐藏字符对应的预测字符；
第二损失生成模块17，用于根据隐藏字符与预测字符生成第二损失；
触发检测生成模块18，用于根据实际论元关系与样本文本语义信息生成触发词检测文本数据；
序列标注模块19，用于预测触发词检测文本数据对应的预测序列标注；
第三损失生成模块20，用于根据触发词检测文本数据对应的实际序列标注与预测序列标注生成第三损失；
模型训练模块21，用于根据第一损失、第二损失及第三损失对初始关系预测模型进行模型参数调整，生成对话关系预测模型；对话关系预测模型用于预测目标对话文本中的目标论元对之间的目标论元关系。
在一种可选的实施方式中，词组预测模块16，包括：
第一确定单元161，用于从所述样本文本语义信息中确定所述隐藏字符对应的掩码隐状态，所述掩码隐状态用于表征所述隐藏字符在所述样本对话文本中对应的语义信息；
字符预测单元162，用于基于所述掩码隐状态，预测所述隐藏字符对应的所述预测字符。
在一种可选的实施方式中，所述装置还包括：
样本获取模块，用于获取原始对话文本和所述样本论元对；
隐藏字符确定模块，用于响应于所述样本论元对中存在至少一个样本论元的论元类型为说话者论元，基于所述样本论元对确定所述隐藏字符；
隐藏处理模块，用于基于所述隐藏字符对所述原始对话文本进行隐藏处理，得到所述样本对话文本。
在一种可选的实施方式中，触发检测生成模块18，包括：
第二确定单元181，用于确定所述实际论元关系对应的论元关系向量；
生成单元182，用于对所述论元关系向量和所述样本文本语义信息进行拼接，生成所述触发词检测文本数据。
在一种可选的实施方式中，所述生成单元182，还用于：
从所述样本文本语义信息中确定出至少一个样本词组隐状态，所述样本词组隐状态用于表征样本对话词组在所述样本对话文本中对应的语义信息；
将所述论元关系向量与所述样本词组隐状态进行拼接，得到所述样本对话词组对应的触发词检测文本；
将各个样本对话词组对应的所述触发词检测文本，确定为所述触发词检测文本数据；
所述序列标注模块19，包括：
触发词预测单元191，用于基于所述触发词检测文本数据进行触发词预测，得到各个所述样本对话词组对应的预测词组标注，所述预测词组标注用于表征所述样本对话词组所属的触发词类型；
第三确定单元192，用于将各个所述样本对话词组对应的所述预测词组标注，确定为所述预测序列标注。
其中，该语义信息提取模块12，包括：
样本拼接单元121，用于在所述初始关系预测模型中基于样本文本拼接符对样本对话文本与样本论元对进行拼接，生成样本拼接文本数据；
论元替换单元122，用于将样本拼接文本数据中的第一样本论元替换为第一样本论元符号，将样本拼接文本数据中的第二样本论元替换为第二样本论元符号，生成样本文本序列数据，所述样本论元对包含所述第一样本论元和所述第二样本论元；
信息提取单元123，用于对样本文本序列数据进行语义特征提取，得到样本对话文本对应的样本文本语义信息。
其中，样本文本拼接符包括样本全局语义符；样本文本序列数据包括样本对话文本对应的样本对话序列数据，样本对话序列数据包括N个样本对话词组，N为正整数；
该信息提取单元123，包括：
隐层提取子单元1231，用于对样本文本序列数据中的样本全局语义符、N个样本对话词组、第一样本论元符号及第二样本论元符号分别进行隐层特征提取，得到样本全局语义符对应的样本全局隐状态、N个样本对话词组分别对应的样本词组隐状态、第一样本论元符号对应的第一初始样本论元隐状态以及第二样本论元符号对应的第二初始样本论元隐状态；
语义确定子单元1232，用于将样本全局隐状态、N个样本词组隐状态、第一初始样本论元隐状态及第二初始样本论元隐状态，确定为样本对话文本对应的样本文本语义信息。
其中，该隐层提取子单元1231，包括：
全局处理子单元123a，用于获取N个样本对话词组、第一样本论元符号及第二样本论元 符号分别与样本全局语义符之间的样本全局关系，对样本全局关系进行特征融合，生成样本全局语义符对应的样本全局隐状态；
状态提取子单元123b，用于对N个样本对话词组、第一样本论元符号及第二样本论元符号分别进行隐层特征提取，得到N个样本对话词组分别对应的样本词组隐状态、第一样本论元符号对应的第一初始样本论元隐状态，以及第二样本论元符号对应的第二初始样本论元隐状态。
其中，N个样本对话词组包括第一样本论元符号及第二样本论元符号；
该关系处理模块13，包括：
第一论元识别单元131，用于从样本文本语义信息所包括的N个样本词组隐状态中，获取第一样本论元符号对应的至少一个第一样本词组隐状态；
对各个第一样本词组隐状态与第一初始样本论元隐状态进行最大池化处理，得到第一样本论元符号对应的第一样本论元隐状态；
第二论元识别单元132，用于从N个样本词组隐状态中获取第二样本论元符号对应的至少一个第二样本词组隐状态；
对各个第二样本词组隐状态与第二初始样本论元隐状态进行最大池化处理，得到第二样本论元符号对应的第二样本论元隐状态；
隐状态拼接单元133，用于对样本全局隐状态、第一样本论元隐状态及第二样本论元隐状态进行拼接，得到样本隐状态信息；
关系预测单元134，用于基于样本隐状态信息预测第一样本论元与第二样本论元之间的预测论元关系。
其中，该关系预测单元134，包括：
语义增强子单元1341，用于对样本隐状态信息进行语义增强，得到样本增强语义信息；
概率选取子单元1342，用于基于样本增强语义信息，确定第一样本论元与第二样本论元对应的M种候选论元关系的样本关系预测概率，M为正整数；
将最大样本关系预测概率对应的候选论元关系，确定为第一样本论元与第二样本论元之间的预测论元关系。
本申请实施例提供了一种对话关系处理装置，通过在初始关系预测模型中增加初始词组预测网络以及初始触发词预测网络，由初始词组预测网络对样本对话文本中被隐藏的说话者论元进行预测，以及由触发词预测网络对样本对话文本中可以指导论元关系的触发词进行预测，以便在损失中引入额外的第二损失和第三损失，对论元关系预测提供辅助信息：说话者特征以及触发词特征，使得初始关系预测模型可以学习到更有利于预测论元关系的有效信息，进而提高论元关系的预测准确性。
进一步地，请参见图10，图10是本申请实施例提供的另一种对话关系处理装置示意图。该对话关系处理装置可以是运行于计算机设备中的一个计算机程序(包括程序代码等)，例如该对话关系处理装置可以为一个应用软件；该装置可以用于执行本申请实施例提供的方法中的相应步骤。如图10所示，该对话关系处理装置1000可以包括：目标模型输入模块31、目标语义获取模块32及目标关系预测模块33。
目标模型输入模块31，用于将目标对话文本及目标论元对输入对话关系预测模型；
目标语义获取模块32，用于在对话关系预测模型中对目标对话文本及目标论元对进行语义特征提取，得到目标对话文本对应的目标文本语义信息，所述目标论元对中的各个目标论元属于所述目标对话文本；
目标关系预测模块33，用于基于对话关系预测模型对目标文本语义信息进行论元关系预测，得到目标论元之间的目标论元关系；对话关系预测模型是采用如上述方面所述的对话关系处理方法训练得到的。
其中，该目标语义获取模块32，包括：
目标拼接单元321，用于基于目标文本拼接符对目标对话文本与目标论元对进行拼接，生成目标拼接文本数据；
目标替换单元322，用于将目标拼接文本数据中的第一目标论元替换为第一目标论元符号，将目标拼接文本数据中的第二目标论元替换为第二目标论元符号，生成目标文本序列数据，所述目标论元对包含所述第一目标论元和所述第二目标论元；
目标获取单元323，用于对目标文本序列数据进行语义特征提取，得到目标对话文本对应的目标文本语义信息。
其中，该目标模型输入模块31，包括：
咨询获取单元311，用于获取目标对话文本与目标对话文本关联的对话咨询信息；
对话解析单元312，用于对对话咨询信息进行解析，提取对话咨询信息所指示的目标论元对；
模型输入单元313，用于将目标对话文本及目标论元对输入对话关系预测模型；
该装置900还包括：
数据存储模块34，用于根据目标对话文本、对话咨询信息及目标论元对生成目标问答数据，将目标问答数据添加至问答数据库中。
参见图11，图11是本申请一个示例性实施例提供的计算机设备的结构示意图。如图11所示，本申请实施例中的计算机设备可以包括：一个或多个处理器1101、存储器1102和输入输出接口1103。该处理器1101、存储器1102和输入输出接口1103通过总线1104连接。存储器1102用于存储计算机程序，该计算机程序包括程序指令，输入输出接口1103用于接收数据及输出数据，如用于模型中各个网络之间进行数据交互，或者用于计算机设备与用户设备之间进行数据交互；处理器1101用于执行存储器1102存储的程序指令。
其中，该处理器1101位于用于进行模型训练的计算机设备时，可以执行如下操作：
将样本对话文本及样本对话文本中的样本论元对输入初始关系预测模型，在初始关系预测模型中对样本对话文本及样本论元对进行特征提取，得到样本对话文本对应的样本文本语义信息；样本论元对包括第一样本论元及第二样本论元；
基于样本文本语义信息预测第一样本论元与第二样本论元之间的预测论元关系，根据第一样本论元与第二样本论元之间的实际论元关系与预测论元关系生成第一损失函数；
获取样本对话文本与样本论元对中的隐藏字符，预测隐藏字符对应的预测字符，根据隐藏字符与预测字符生成第二损失函数；
根据实际论元关系与样本文本语义信息生成触发词检测文本数据，预测触发词检测文本数据中的预测序列标注，根据触发词检测文本数据中的实际序列标注与预测序列标注生成第三损失函数；
根据第一损失函数、第二损失函数及第三损失函数对初始关系预测模型进行模型参数调整，生成对话关系预测模型；对话关系预测模型用于预测目标对话文本中的目标论元对之间的目标论元关系。
其中，该处理器1101位于用于进行模型预测的计算机设备时，可以执行如下操作：
将目标对话文本及目标对话文本中的目标论元对输入对话关系预测模型，在对话关系预测模型中对目标对话文本及目标论元对进行特征提取，得到目标对话文本对应的目标文本语义信息；目标论元对包括第一目标论元及第二目标论元；
基于对话关系预测模型中的关系预测网络对目标文本语义信息进行预测，得到第一目标论元与第二目标论元之间的目标论元关系。
在一些可行的实施方式中，该处理器1101可以是中央处理单元(Central Processing Unit，CPU)，该处理器还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑 器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
该存储器1102可以包括只读存储器和随机存取存储器，并向处理器1101和输入输出接口1103提供指令和数据。存储器1102的一部分还可以包括非易失性随机存取存储器。例如，存储器1102还可以存储设备类型的信息。
具体实现中，该计算机设备可通过其内置的各个功能模块执行上述各个方法实施例所提供的实现方式。
本申请实施例通过提供一种计算机设备，包括：处理器、输入输出接口、存储器，通过处理器获取存储器中的计算机程序，执行该图4中所示方法的各个步骤，进行对话关系处理操作。本申请实施例通过在初始关系预测模型中增加初始词组预测网络以及初始触发词预测网络，由初始词组预测网络对样本对话文本中被隐藏的说话者论元进行预测，以及由触发词预测网络对样本对话文本中可以指导论元关系的触发词进行预测，以便在损失中引入额外的第二损失和第三损失，对论元关系预测提供辅助信息：说话者特征以及触发词特征，使得初始关系预测模型可以学习到更有利于预测论元关系的有效信息，进而提高论元关系的预测准确性。
本申请实施例还提供一种计算机可读存储介质，该计算机可读存储介质存储有计算机程序，该计算机程序适于由该处理器加载并执行图4或图8中各个步骤所提供的对话关系处理方法，具体可参见该图4或图8中各个步骤所提供的实现方式，在此不再赘述。另外，对采用相同方法的有益效果描述，也不再进行赘述。对于本申请所涉及的计算机可读存储介质实施例中未披露的技术细节，请参照本申请方法实施例的描述。作为示例，计算机程序可被部署为在一个计算机设备上执行，或者在位于一个地点的多个计算机设备上执行，又或者，在分布在多个地点且通过通信网络互连的多个计算机设备上执行。
该计算机可读存储介质可以是前述任一实施例提供的对话关系处理装置或者该计算机设备的内部存储单元，例如计算机设备的硬盘或内存。该计算机可读存储介质也可以是该计算机设备的外部存储设备，例如该计算机设备上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(secure digital，SD)卡，闪存卡(flash card)等。进一步地，该计算机可读存储介质还可以既包括该计算机设备的内部存储单元也包括外部存储设备。该计算机可读存储介质用于存储该计算机程序以及该计算机设备所需的其他程序和数据。该计算机可读存储介质还可以用于暂时地存储已经输出或者将要输出的数据。
本申请实施例还提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行图4或图8中的各种可选方式中所提供的方法，实现了将对话式关系抽取分为三个相关的子任务，分别是关系预测任务、词组预测任务及触发词预测任务，通过综合这三个子任务，对模型进行联合训练，可以充分利用从词组预测任务及触发词预测任务中学到的有效信息，并基于该有效信息影响关系预测任务，从而提高对话关系处理的准确性。
本申请实施例的说明书和权利要求书及附图中的术语“第一”、“第二”等是用于区别不同对象，而非用于描述特定顺序。此外，术语“包括”以及它们任何变形，意图在于覆盖不排他的包含。例如包含了一系列步骤或单元的过程、方法、装置、产品或设备没有限定于已列出的步骤或模块，而是可选地还包括没有列出的步骤或模块，或可选地还包括对于这些过程、方法、装置、产品或设备固有的其他步骤单元。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，为了清楚地说明硬件和软件的可互换性，在该说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可 以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
本申请实施例提供的方法及相关装置是参照本申请实施例提供的方法流程图和/或结构示意图来描述的，具体可由计算机程序指令实现方法流程图和/或结构示意图的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。这些计算机程序指令可提供到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或结构示意图一个方框或多个方框中指定的功能的装置。这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或结构示意图一个方框或多个方框中指定的功能。这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或结构示意一个方框或多个方框中指定的功能的步骤。
本申请实施例方法中的步骤可以根据实际需要进行顺序调整、合并和删减。
本申请实施例装置中的模块可以根据实际需要进行合并、划分和删减。
以上所揭露的仅为本申请可选实施例而已，当然不能以此来限定本申请之权利范围，因此依本申请权利要求所作的等同变化，仍属本申请所涵盖的范围。

Claims (16)

  1. 一种对话关系处理方法，其特征在于，所述方法应用于计算机设备，所述方法包括：
    通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息，所述样本论元对中的各个样本论元属于所述样本对话文本；
    基于所述样本文本语义信息进行论元关系预测，并基于论元关系预测结果确定第一损失，所述论元关系预测结果用于表征所述样本论元之间的关系；
    基于所述样本文本语义信息进行隐藏字符预测，并基于隐藏字符预测结果确定第二损失，所述隐藏字符预测结果用于表征所述样本对话文本中的隐藏字符；
    基于所述样本文本语义信息和实际论元关系进行触发词预测，并基于触发词预测结果确定第三损失，所述实际论元关系为所述样本论元对应的标注论元关系，所述触发词预测结果用于表征触发词在所述样本对话文本中的位置；
    基于所述第一损失、所述第二损失和所述第三损失，训练所述初始关系预测模型，得到对话关系预测模型。
  2. 根据权利要求1所述的方法，其特征在于，
    所述通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息包括：
    将所述样本对话文本及所述样本论元对输入所述初始关系预测模型，在所述初始关系预测模型中对所述样本对话文本及所述样本论元对进行语义特征提取，得到所述样本文本语义信息；
    所述基于所述样本文本语义信息进行论元关系预测，并基于论元关系预测结果确定第一损失，包括：
    基于所述样本文本语义信息预测所述样本论元之间的预测论元关系；
    根据所述样本论元之间的所述实际论元关系与所述预测论元关系生成所述第一损失；
    所述基于所述样本文本语义信息进行隐藏字符预测，并基于隐藏字符预测结果确定第二损失，包括：
    获取所述样本对话文本中的所述隐藏字符；
    基于所述样本文本语义信息，预测所述隐藏字符对应的预测字符；
    根据所述隐藏字符与所述预测字符生成所述第二损失；
    所述基于所述样本文本语义信息和实际论元关系进行触发词预测，并基于触发词预测结果确定第三损失，包括：
    根据所述实际论元关系与所述样本文本语义信息生成触发词检测文本数据；
    预测所述触发词检测文本数据对应的所述预测序列标注；
    根据所述触发词检测文本数据对应的实际序列标注与所述预测序列标注生成所述第三损失；
    所述基于所述第一损失、所述第二损失和所述第三损失，训练所述初始关系预测模型，得到关系预测模型，包括：
    根据所述第一损失、所述第二损失及所述第三损失对所述初始关系预测模型进行模型参数调整，生成对话关系预测模型；所述对话关系预测模型用于预测目标对话文本中的目标论元对对应的目标论元关系。
  3. 根据权利要求2所述的方法，其特征在于，所述基于所述样本文本语义信息，预测所述隐藏字符对应的预测字符，包括：
    从所述样本文本语义信息中确定所述隐藏字符对应的掩码隐状态，所述掩码隐状态用于表征所述隐藏字符在所述样本对话文本中对应的语义信息；
    基于所述掩码隐状态，预测所述隐藏字符对应的所述预测字符。
  4. 根据权利要求3所述的方法，其特征在于，所述通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息之前，所述方法还包括：
    获取原始对话文本和所述样本论元对；
    响应于所述样本论元对中存在至少一个样本论元的论元类型为说话者论元，基于所述样本论元对确定所述隐藏字符；
    基于所述隐藏字符对所述原始对话文本进行隐藏处理，得到所述样本对话文本。
  5. 根据权利要求2所述的方法，其特征在于，所述根据所述实际论元关系与所述样本文本语义信息生成触发词检测文本数据，包括：
    确定所述实际论元关系对应的论元关系向量；
    对所述论元关系向量和所述样本文本语义信息进行拼接，生成所述触发词检测文本数据。
  6. 根据权利要求5所述的方法，其特征在于，所述对所述论元关系向量和所述样本文本语义信息进行拼接，生成所述触发词检测文本数据，包括：
    从所述样本文本语义信息中确定出至少一个样本词组隐状态，所述样本词组隐状态用于表征样本对话词组在所述样本对话文本中对应的语义信息；
    将所述论元关系向量与所述样本词组隐状态进行拼接，得到所述样本对话词组对应的触发词检测文本；
    将各个样本对话词组对应的所述触发词检测文本，确定为所述触发词检测文本数据；
    所述预测所述触发词检测文本数据对应的所述预测序列标注，包括：
    基于所述触发词检测文本数据进行触发词预测，得到各个所述样本对话词组对应的预测词组标注，所述预测词组标注用于表征所述样本对话词组所属的触发词类型；
    将各个所述样本对话词组对应的所述预测词组标注，确定为所述预测序列标注。
  7. 根据权利要求1至5任一所述的方法，其特征在于，所述通过初始关系预测模型对样本对话文本和样本论元对进行语义特征提取，得到样本文本语义信息，包括：
    在所述初始关系预测模型中基于样本文本拼接符对所述样本对话文本与所述样本论元对进行拼接，生成样本拼接文本数据；
    将所述样本拼接文本数据中的所述第一样本论元替换为第一样本论元符号，将所述样本拼接文本数据中的所述第二样本论元替换为第二样本论元符号，生成样本文本序列数据，所述样本论元对包含所述第一样本论元和所述第二样本论元；
    对所述样本文本序列数据进行语义特征提取，得到所述样本文本语义信息。
  8. 根据权利要求7所述的方法，其特征在于，所述样本文本拼接符包括样本全局语义符；所述样本文本序列数据包括所述样本对话文本对应的样本对话序列数据，所述样本对话序列数据包括N个样本对话词组，N为正整数；
    所述对所述样本文本序列数据进行语义特征提取，得到所述样本文本语义信息，包括：
    对所述样本文本序列数据中的所述样本全局语义符、所述N个样本对话词组、所述第一样本论元符号及所述第二样本论元符号分别进行隐层特征提取，得到所述样本全局语义符对应的样本全局隐状态、所述N个样本对话词组分别对应的样本词组隐状态、所述第一样本论元符号对应的第一初始样本论元隐状态以及所述第二样本论元符号对应的第二初始样本论元隐状态；
    将所述样本全局隐状态、N个样本词组隐状态、所述第一初始样本论元隐状态及所述第二初始样本论元隐状态，确定为所述样本对话文本对应的样本文本语义信息。
  9. 根据权利要求8所述的方法，其特征在于，所述对所述样本文本序列数据中的所述样本全局语义符、所述N个样本对话词组、所述第一样本论元符号及所述第二样本论元符号分别进行隐层特征提取，得到所述样本全局语义符对应的样本全局隐状态、所述N个样本对话词组分别对应的样本词组隐状态、所述第一样本论元符号对应的第一初始样本论元隐状态以及所述第二样本论元符号对应的第二初始样本论元隐状态，包括：
    获取所述N个样本对话词组、所述第一样本论元符号及所述第二样本论元符号分别与所述样本全局语义符之间的样本全局关系，对所述样本全局关系进行特征融合，生成所述样本全局语义符对应的所述样本全局隐状态；
    对所述N个样本对话词组、所述第一样本论元符号及所述第二样本论元符号分别进行隐层特征提取，得到所述N个样本对话词组分别对应的所述样本词组隐状态、所述第一样本论元符号对应的所述第一初始样本论元隐状态，以及所述第二样本论元符号对应的所述第二初始样本论元隐状态。
  10. 根据权利要求8所述的方法，其特征在于，所述N个样本对话词组包括所述第一样本论元符号及所述第二样本论元符号；
    所述基于所述样本文本语义信息进行论元关系预测，包括：
    从所述样本文本语义信息所包括的所述N个样本词组隐状态中，获取所述第一样本论元符号对应的至少一个第一样本词组隐状态；
    对各个所述第一样本词组隐状态与所述第一初始样本论元隐状态进行最大池化处理，得到所述第一样本论元符号对应的第一样本论元隐状态；
    从所述N个样本词组隐状态中获取所述第二样本论元符号对应的至少一个第二样本词组隐状态；
    对各个所述第二样本词组隐状态与所述第二初始样本论元隐状态进行最大池化处理，得到所述第二样本论元符号对应的第二样本论元隐状态；
    对所述样本全局隐状态、所述第一样本论元隐状态及所述第二样本论元隐状态进行拼接，得到样本隐状态信息；
    基于所述样本隐状态信息预测所述第一样本论元与所述第二样本论元之间的预测论元关系。
  11. 根据权利要求10所述的方法，其特征在于，所述基于所述样本隐状态信息预测所述第一样本论元与所述第二样本论元之间的预测论元关系，包括：
    对所述样本隐状态信息进行语义增强，得到样本增强语义信息；
    基于所述样本增强语义信息，确定所述第一样本论元与所述第二样本论元对应的M种候选论元关系的样本关系预测概率，M为正整数；
    将最大样本关系预测概率对应的候选论元关系，确定为所述第一样本论元与所述第二样本论元之间的所述预测论元关系。
  12. 一种对话关系处理方法，其特征在于，所述方法应用于计算机设备，所述方法包括：
    将目标对话文本及目标论元对输入对话关系预测模型，在所述对话关系预测模型中对所述目标对话文本及所述目标论元对进行语义特征提取，得到所述目标对话文本对应的目标文本语义信息，所述目标论元对中的各个目标论元属于所述目标对话文本；
    基于对话关系预测模型对所述目标文本语义信息进行论元关系预测，得到所述目标论元之间的目标论元关系；所述对话关系预测模型是采用如权利要求1所述的对话关系处理方法训练得到的。
  13. 根据权利要求12所述的方法，其特征在于，所述在所述对话关系预测模型中对所述目标对话文本及所述目标论元对进行语义特征提取，得到所述目标对话文本对应的目标文本语义信息，包括：
    基于目标文本拼接符对所述目标对话文本与所述目标论元对进行拼接，生成目标拼接文本数据；
    将所述目标拼接文本数据中的第一目标论元替换为第一目标论元符号，将所述目标拼接文本数据中的第二目标论元替换为第二目标论元符号，生成目标文本序列数据，所述目标论元对包含所述第一目标论元和所述第二目标论元；
    对所述目标文本序列数据进行语义特征提取，得到所述目标对话文本对应的所述目标文 本语义信息。
  14. 根据权利要求12所述的方法，其特征在于，所述将目标对话文本及目标论元对输入对话关系预测模型，包括：
    获取目标对话文本与所述目标对话文本关联的对话咨询信息；
    对所述对话咨询信息进行解析，提取所述对话咨询信息所指示的所述目标论元对；
    将所述目标对话文本及所述目标论元对输入所述对话关系预测模型；
    所述方法还包括：
    根据所述目标对话文本、所述对话咨询信息及所述目标论元对生成目标问答数据，将所述目标问答数据添加至问答数据库中。
  15. 一种计算机设备，其特征在于，包括处理器、存储器、输入输出接口；
    所述处理器分别与所述存储器和所述输入输出接口相连，其中，所述输入输出接口用于接收数据及输出数据，所述存储器用于存储计算机程序，所述处理器用于调用所述计算机程序，以使得所述计算机设备执行权利要求1-11任一项所述的对话关系处理方法，或者执行权利要求12-14任一项所述的对话关系处理方法。
  16. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有计算机程序，所述计算机程序适于由处理器加载并执行，以使得具有所述处理器的计算机设备执行权利要求1-11任一项所述的对话关系处理方法，或者执行权利要求12-14任一项所述的对话关系处理方法。
PCT/CN2021/108503 2021-06-17 2021-07-26 一种对话关系处理方法、计算机及可读存储介质 WO2022262080A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/132,723 US20230244878A1 (en) 2021-06-17 2023-04-10 Extracting conversational relationships based on speaker prediction and trigger word prediction

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110674476.3A CN114676705B (zh) 2021-06-17 2021-06-17 一种对话关系处理方法、计算机及可读存储介质
CN202110674476.3 2021-06-17

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/132,723 Continuation US20230244878A1 (en) 2021-06-17 2023-04-10 Extracting conversational relationships based on speaker prediction and trigger word prediction

Publications (1)

Publication Number Publication Date
WO2022262080A1 true WO2022262080A1 (zh) 2022-12-22

Family

ID=82070946

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/108503 WO2022262080A1 (zh) 2021-06-17 2021-07-26 一种对话关系处理方法、计算机及可读存储介质

Country Status (3)

Country Link
US (1) US20230244878A1 (zh)
CN (1) CN114676705B (zh)
WO (1) WO2022262080A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116127328A (zh) * 2023-04-13 2023-05-16 北京轮子科技有限公司 一种对话状态识别模型的训练方法、装置、介质及设备

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2024207568A1 (zh) * 2023-04-04 2024-10-10 浙江棱镜全息科技有限公司 一种用于车载智能交互的中间gpt模型构建方法和装置

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140372102A1 (en) * 2013-06-18 2014-12-18 Xerox Corporation Combining temporal processing and textual entailment to detect temporally anchored events
US20180349326A1 (en) * 2017-05-30 2018-12-06 International Business Machines Corporation Weather-based natural language text processing
CN111651606A (zh) * 2020-06-05 2020-09-11 深圳市慧择时代科技有限公司 一种文本处理方法、装置及电子设备
CN112069811A (zh) * 2020-08-24 2020-12-11 武汉大学 多任务交互增强的电子文本事件抽取方法
CN112116075A (zh) * 2020-09-18 2020-12-22 厦门安胜网络科技有限公司 事件提取模型生成方法和装置、文本事件提取方法和装置
CN112163416A (zh) * 2020-10-09 2021-01-01 北京理工大学 一种融合句法和实体关系图卷积网络的事件联合抽取方法
CN112580328A (zh) * 2020-12-11 2021-03-30 上海明略人工智能(集团)有限公司 事件信息的抽取方法及装置、存储介质、电子设备
CN112860852A (zh) * 2021-01-26 2021-05-28 北京金堤科技有限公司 信息分析方法及装置、电子设备和计算机可读存储介质
CN112861527A (zh) * 2021-03-17 2021-05-28 合肥讯飞数码科技有限公司 一种事件抽取方法、装置、设备及存储介质
CN112949307A (zh) * 2021-02-25 2021-06-11 平安科技（深圳）有限公司 预测语句实体的方法、装置和计算机设备
CN112966079A (zh) * 2021-03-02 2021-06-15 中国电子科技集团公司第二十八研究所 一种用于对话系统的面向事件画像的文本分析方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8155946B2 (en) * 2002-12-23 2012-04-10 Definiens Ag Computerized method and system for searching for text passages in text documents
US11494547B2 (en) * 2016-04-13 2022-11-08 Microsoft Technology Licensing, Llc Inputting images to electronic devices
CN111488739B (zh) * 2020-03-17 2023-07-18 天津大学 基于多粒度生成图像增强表示的隐式篇章关系识别方法

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140372102A1 (en) * 2013-06-18 2014-12-18 Xerox Corporation Combining temporal processing and textual entailment to detect temporally anchored events
US20180349326A1 (en) * 2017-05-30 2018-12-06 International Business Machines Corporation Weather-based natural language text processing
CN111651606A (zh) * 2020-06-05 2020-09-11 深圳市慧择时代科技有限公司 一种文本处理方法、装置及电子设备
CN112069811A (zh) * 2020-08-24 2020-12-11 武汉大学 多任务交互增强的电子文本事件抽取方法
CN112116075A (zh) * 2020-09-18 2020-12-22 厦门安胜网络科技有限公司 事件提取模型生成方法和装置、文本事件提取方法和装置
CN112163416A (zh) * 2020-10-09 2021-01-01 北京理工大学 一种融合句法和实体关系图卷积网络的事件联合抽取方法
CN112580328A (zh) * 2020-12-11 2021-03-30 上海明略人工智能(集团)有限公司 事件信息的抽取方法及装置、存储介质、电子设备
CN112860852A (zh) * 2021-01-26 2021-05-28 北京金堤科技有限公司 信息分析方法及装置、电子设备和计算机可读存储介质
CN112949307A (zh) * 2021-02-25 2021-06-11 平安科技（深圳）有限公司 预测语句实体的方法、装置和计算机设备
CN112966079A (zh) * 2021-03-02 2021-06-15 中国电子科技集团公司第二十八研究所 一种用于对话系统的面向事件画像的文本分析方法
CN112861527A (zh) * 2021-03-17 2021-05-28 合肥讯飞数码科技有限公司 一种事件抽取方法、装置、设备及存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116127328A (zh) * 2023-04-13 2023-05-16 北京轮子科技有限公司 一种对话状态识别模型的训练方法、装置、介质及设备
CN116127328B (zh) * 2023-04-13 2023-08-25 北京轮子科技有限公司 一种对话状态识别模型的训练方法、装置、介质及设备

Also Published As

Publication number Publication date
CN114676705B (zh) 2024-10-15
US20230244878A1 (en) 2023-08-03
CN114676705A (zh) 2022-06-28

Similar Documents

Publication Publication Date Title
CN111104498A (zh) 一种任务型对话系统中的语义理解方法
CN111831826B (zh) 跨领域的文本分类模型的训练方法、分类方法以及装置
US11531821B2 (en) Intent resolution for chatbot conversations with negation and coreferences
US20230244878A1 (en) Extracting conversational relationships based on speaker prediction and trigger word prediction
CN112287095A (zh) 确定问题答案的方法、装置、计算机设备及存储介质
US20220414463A1 (en) Automated troubleshooter
CN113158656B (zh) 讽刺内容识别方法、装置、电子设备以及存储介质
WO2023045184A1 (zh) 一种文本类别识别方法、装置、计算机设备及介质
CN111368066B (zh) 获取对话摘要的方法、装置和计算机可读存储介质
CN113051895A (zh) 语音识别的方法、装置、电子设备、介质和程序产品
CN110647613A (zh) 一种课件构建方法、装置、服务器和存储介质
CN115115984A (zh) 视频数据处理方法、装置、程序产品、计算机设备和介质
CN114065720A (zh) 会议纪要生成方法、装置、存储介质及电子设备
CN114925206A (zh) 人工智能体、语音信息识别方法、存储介质和程序产品
CN113889092A (zh) 语音识别结果的后处理模型的训练方法、处理方法及装置
KR102434666B1 (ko) 사전 데이터베이스를 활용하여 음성 데이터에 기반한 텍스트를 생성하기 위한 방법 및 컴퓨팅 장치
CN112233648A (zh) 结合rpa及ai的数据的处理方法、装置、设备及存储介质
CN115115432B (zh) 基于人工智能的产品信息推荐方法及装置
CN113392190B (zh) 一种文本识别方法、相关设备及装置
CN116186219A (zh) 一种人机对话交互方法方法、系统及存储介质
CN111401069A (zh) 会话文本的意图识别方法、意图识别装置及终端
KR102624790B1 (ko) 다중 의도 발화의 의도 분석 및 처리를 위한 자연어 처리 장치, 프로그램 및 그것의 제어 방법
CN115623134A (zh) 会议音频处理方法、装置、设备及存储介质
CN112749553B (zh) 视频文件的文本信息处理方法、装置和服务器
WO2023272833A1 (zh) 一种数据检测方法、装置、设备及可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21945661

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 03/04/2024)