WO2022142028A1 - 对话状态确定方法、终端设备及存储介质 - Google Patents

对话状态确定方法、终端设备及存储介质 Download PDF

Info

Publication number
WO2022142028A1
WO2022142028A1 PCT/CN2021/091132 CN2021091132W WO2022142028A1 WO 2022142028 A1 WO2022142028 A1 WO 2022142028A1 CN 2021091132 W CN2021091132 W CN 2021091132W WO 2022142028 A1 WO2022142028 A1 WO 2022142028A1
Authority
WO
WIPO (PCT)
Prior art keywords
information
current
slot
candidate
domain
Prior art date
Application number
PCT/CN2021/091132
Other languages
English (en)
French (fr)
Inventor
陈海滨
Original Assignee
平安科技（深圳）有限公司
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 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2022142028A1 publication Critical patent/WO2022142028A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/332—Query formulation
    • G06F16/3329—Natural language query formulation or dialogue systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3344—Query execution using natural language analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3347—Query execution using vector based model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present application relates to the technical field of artificial intelligence, and in particular, to a method for determining a dialog state, a terminal device and a storage medium.
  • a question-and-answer between a user and a smart device is usually called a round of dialogue.
  • smart devices can usually obtain necessary information through multiple rounds of dialogue to finally obtain clear user instructions.
  • multiple rounds of dialogue usually correspond to the execution of a task.
  • necessary information for executing the task of purchasing an airline ticket can be obtained through multiple rounds of dialogue, so as to obtain a clear user instruction based on the obtained necessary information, so as to accurately execute the task of purchasing an airline ticket.
  • the inventor realizes that it is necessary to obtain the necessary information of the task to be executed through multiple rounds of dialogues, and in the related art, the efficiency of obtaining the necessary information of the task to be executed is not high enough.
  • the present application aims to provide a dialog state determination method, a terminal device and a storage medium, which can improve the efficiency of obtaining necessary information of a task to be performed to a certain extent.
  • a method for determining a dialog state including:
  • a device for determining a dialog state including:
  • a sentence receiving unit configured to obtain a system response sentence for the previous input sentence in response to receiving the current input sentence input by the user in the current round of dialogue, wherein the previous input sentence is the sentence input by the user in the previous round of dialogue ;
  • the information selection unit is used to select the candidate field information that matches the current input sentence and the system response sentence from the pre-built candidate field information set as the current field information, and from the pre-built candidate slot information set, select the candidate field information set that matches the current input sentence.
  • the candidate slot information matching the statement and the system response statement is used as the current slot information;
  • the information determination unit is used to input the current domain information, historical domain status information, current slot information and historical slot status information into the pre-trained dialogue state model to obtain the dialogue state information of the current round of dialogue, wherein the dialogue state information includes Current realm status information and current slot status information.
  • a terminal device including a memory, a processor, and a computer program stored in the memory and running on the terminal device, where the processor implements the following steps when executing the computer program:
  • a computer-readable storage medium which stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:
  • the beneficial effect of the embodiment of the present application is: by introducing domain factors, information in different domains can be distinguished, and necessary information corresponding to tasks in different domains can be collected in one multi-round dialogue.
  • the present application can save computing resources and improve the efficiency of obtaining necessary information of tasks to be performed. It helps to improve the efficiency of human-machine dialogue.
  • Fig. 1 is the realization flow chart of a kind of dialog state determination method provided by the embodiment of the present application
  • Fig. 2 is the realization flow chart of the construction method of the candidate domain information set provided by the embodiment of the present application;
  • FIG. 3 is an implementation flowchart of a method for determining a dialog state provided by another embodiment of the present application.
  • Fig. 4 is the realization flow chart of a dialog state determination method provided by still another embodiment of the present application.
  • FIG. 5 is a structural block diagram of an apparatus for determining a dialog state provided by an embodiment of the present application.
  • FIG. 6 is a structural block diagram of a terminal device provided by an embodiment of the present application.
  • the dialog state determination method involved in the embodiment of the present application may be executed by a control device or a terminal (hereinafter referred to as "terminal device").
  • FIG. 1 shows an implementation flowchart of a method for determining a dialog state provided by an embodiment of the present application, including:
  • Step 101 in response to receiving the current input sentence input by the user in the current round of dialogue, obtain a system response sentence for the previous input sentence.
  • the current input sentence is usually the sentence input by the user in the current round of dialogue.
  • the last input sentence usually the sentence entered by the user in the previous round of dialogue.
  • the system response sentence is usually the sentence that the terminal device responds to the sentence input by the user in the previous round of dialogue.
  • the system responds to the sentence, and one input and one response are a round of dialogue.
  • the above-mentioned executive body may receive the current input sentence input by the user, and upon receiving the current input sentence, call up the system response sentence that the executive body responded to the last input sentence input by the user in the previous round of dialogue.
  • the above-mentioned current input sentence and the above-mentioned system response sentence can usually be in the form of speech or text.
  • the execution body may first convert the speech into text, and then convert the text into a vector that is convenient for data processing. If the current input sentence is in the form of text, the execution body can directly convert the text into a vector that is convenient for data processing.
  • the executive body can use a bidirectional long short-term memory network (Bi-directional Long Short-Term Memory, Bi-LSTM) converts the text corresponding to the sentence into a vector.
  • Bi-LSTM bidirectional Long Short-Term Memory
  • the execution body can also use the word2vec model to convert the words in the sentence into vectors, and then splicing the vectors corresponding to each word to form a vector whose dimension is the sum of the dimensions of all word vectors, and obtain the vector corresponding to the sentence.
  • the word2vec model is used to represent the correspondence between words in a sentence and word vectors
  • the word2vec model is a neural network model commonly used by those skilled in the art to convert words into word vectors, which will not be repeated here.
  • Step 102 From the pre-built candidate domain information set, select the candidate domain information that matches the current input sentence and the system response sentence as the current domain information, and from the pre-built candidate slot information set, select the current input sentence and system.
  • the candidate slot information matched by the response statement is used as the current slot information.
  • the domain information is usually the information used to indicate the target task.
  • the target task can be various tasks set in advance.
  • a task could be a loan, repayment, buying a plane ticket, etc.
  • the candidate domain information is generally domain information for candidates.
  • the slot information usually includes a slot and a slot value.
  • the slot usually refers to the key information that needs to be collected from the user.
  • a slot usually has a slot value, and the slot value is usually a specific value of the key information expressed by the user.
  • the slot may be the price, and the slot value may be the value of the price.
  • the slot may be the departure place and the slot value may be Shanghai.
  • slot filling is usually the process of collecting user information to fill in ambiguous or missing user intent.
  • the above-mentioned execution body can obtain the current domain information and the current slot information in the following manner: from the candidate domain information set, select the candidate domain information that has the most repeated words with the current input sentence and the system response sentence as the current domain information. And from the candidate slot information set, the candidate slot information with the most repeated words with the current input sentence and the system response sentence is selected as the current slot information.
  • Step 103 Input the current domain information, historical domain state information, current slot information and historical slot state information into the pre-trained dialogue state model to obtain the dialogue state information of the current round of dialogue.
  • the dialog state information includes current domain state information and current slot state information.
  • the current domain state information is usually the domain information corresponding to the current round of dialogue and each round of dialogues before the current round of dialogue.
  • the current field status information may include multiple pieces of information for indicating target tasks, that is, the current field status information may simultaneously indicate multiple target tasks.
  • the current slot status information is usually the slot information corresponding to the current round of dialogue and the comprehensive corresponding to the rounds of dialogues before the current round of dialogue.
  • the historical domain state information is usually the domain information corresponding to the synthesis of each round of dialogue before the current round of dialogue.
  • the historical slot status information is usually the slot information corresponding to the synthesis of each round of dialogue before the current round of dialogue.
  • the dialogue state model is usually used to represent current domain information, historical domain state information, current slot information, and the correspondence between historical slot state information and the current dialogue state.
  • the dialogue state model may be generated based on statistics of a large number of states of a large number of dialogues, and stores a plurality of current domain information, historical domain state information, current slot information, and the correspondence between historical slot state information and dialogue state information
  • the corresponding relationship table of the relationship can also be based on the training samples, using machine learning methods, for the initial model (such as convolutional neural network Neural Network, CNN), residual network (ResNet, etc.) after training.
  • the initial model such as convolutional neural network Neural Network, CNN), residual network (ResNet, etc.
  • the present application can save computing resources and improve the efficiency of obtaining necessary information of tasks to be performed. It helps to improve the efficiency of human-machine dialogue.
  • the dialogue status information is obtained based on the comprehensive analysis of the current round of dialogue and the rounds of dialogues before the current round of dialogue. For each round of dialogue, the corresponding dialogue status information can be obtained.
  • the dialog state information of the task can improve the efficiency of obtaining necessary information of the task to be executed. It helps to improve the efficiency of human-machine dialogue.
  • the above-mentioned execution body may also obtain the current domain information and the current slot information in the following manner: First, from the candidate domain information set, select the candidate domain that is most similar to the semantics of the combined statement information, as current domain information. Then, from the candidate slot information set, the candidate slot information closest to the semantics of the combined sentence is selected as the current slot information.
  • the combined statement may be a statement generated by combining the current input statement and the system response statement. For example, if the current input sentence is: Help to buy a ticket to Shenzhen, the system response sentence for the previous round of input is: What can I do for help? Then the combined statement can be: Need my help to do something, help to buy a ticket to Shenzhen.
  • the above-mentioned executive body may use a semantic similarity algorithm, such as a deep semantic model algorithm (Deep Structured Semantic Model). models, DSSM), convolutional latent semantic model (Convolutional Latent Semantic Model, CLSM), etc., calculate the similarity between the combined sentence and each candidate domain information, and then select the candidate domain information with the largest corresponding similarity as the current domain information. and using the above semantic similarity algorithm to calculate the similarity between the combined sentence and each candidate slot information, and then select the candidate slot information with the largest corresponding similarity as the current slot information.
  • a semantic similarity algorithm such as a deep semantic model algorithm (Deep Structured Semantic Model). models, DSSM), convolutional latent semantic model (Convolutional Latent Semantic Model, CLSM), etc.
  • the candidate domain information with the most similar semantics is selected as the current domain information
  • the candidate slot information with the most similar semantics is selected as the current slot information, which can more accurately capture the real intention of the user, thereby further improving the human-computer dialogue. s efficiency.
  • selecting the candidate domain information that is most similar to the semantics of the combined sentence from the candidate domain information set as the current domain information including: first, for the candidate domain information set in the candidate domain information set information, and determine the semantic similarity between the vector corresponding to the candidate domain information and the vector corresponding to the combined sentence. Then, from the candidate domain information set, the candidate domain information with the largest corresponding semantic similarity is selected as the current domain information.
  • semantic similarity is usually a measure used to describe the semantic correlation between data.
  • the value of semantic similarity is usually greater than 0 and less than or equal to 1.
  • the semantic similarity may include any one or more of the following: cosine similarity, reciprocal of Euclidean distance, and the like.
  • the above-mentioned executive body may calculate the semantic similarity between the vector of each candidate domain information and the vector of the combined sentence, and then select the candidate domain information with the largest corresponding semantic similarity as the current domain information.
  • the similarity between the two vectors is directly used as the semantic similarity between the candidate domain information and the combined sentence.
  • the computational complexity is low, and the efficiency of obtaining the necessary information for the task to be performed can be guaranteed.
  • computing resources are saved.
  • the candidate slot information that is most similar to the semantics of the combined sentence is selected as the current slot information, including: for the candidates in the candidate slot information set Slot information, to determine the semantic similarity between the vector corresponding to the candidate slot information and the vector corresponding to the combined sentence. From the candidate slot information set, the candidate slot information with the largest corresponding semantic similarity is selected as the current slot information.
  • the above-mentioned executive body may calculate the semantic similarity between the vector of each candidate slot information and the vector of the combined sentence, and then select the candidate slot information with the largest corresponding semantic similarity as the current slot information.
  • the similarity between the two vectors is directly used as the semantic similarity between the candidate slot information and the combined statement.
  • the computational complexity is low, and the efficiency of obtaining the necessary information for the task to be performed can be guaranteed. At the same time, it further saves computing resources.
  • the dialogue state model is obtained by training through the following steps: first, a training sample set is obtained, and the training samples in the training sample set include domain information, historical domain state information, slot information, and historical slot information. Bit state information and corresponding dialog state information. Then, the domain information, historical domain status information, slot information, and historical slot status information of the training samples in the training sample set are used as input, and the input domain information, historical domain status information, slot information, and historical slot status The dialogue state information corresponding to the information is used as the expected output, and the dialogue state model is obtained by training.
  • the dialog state information includes current domain state information and current slot state information.
  • the current domain state information is usually the domain information corresponding to the current round of dialogue and each round of dialogues before the current round of dialogue.
  • the current field status information may include multiple pieces of information for indicating target tasks, that is, the current field status information may simultaneously indicate multiple target tasks.
  • the current slot status information is usually the slot information corresponding to the current round of dialogue and the comprehensive corresponding to the rounds of dialogues before the current round of dialogue.
  • the historical domain state information is usually the domain information corresponding to the synthesis of each round of dialogue before the current round of dialogue.
  • the historical slot status information is usually the slot information corresponding to the synthesis of each round of dialogue before the current round of dialogue.
  • the dialogue state model obtained by training can output the current domain state information and the current slot of the current round of dialogue according to the input current domain information, historical domain state information, current slot information and historical slot state information status information. Only one model needs to be trained to obtain two kinds of information at the same time, namely, the current domain state information and the current slot state information. Compared with the related art, at least two models need to be trained to obtain the current domain state information and the current slot state information respectively, the present application can speed up the model training efficiency.
  • FIG. 2 is an implementation flowchart of a method for constructing a candidate domain information set provided by another embodiment of the present application. Details are as follows:
  • Step 201 domain calculation step: determine the domain similarity between the target domain training sentence and the initial domain information, and in response to the domain similarity being greater than or equal to a preset domain similarity threshold, determine the initial domain information as candidate domain information.
  • the target domain training sentence includes the information of the task indicated by the candidate domain information.
  • the training sentence in the target domain can be: Help to buy a plane ticket to Shenzhen.
  • the task indicated by the candidate domain information is: buy an air ticket.
  • domain similarity is usually a numerical value used to describe the similarity of the tasks pointed to by two pieces of information.
  • the value of domain similarity is usually greater than 0 and less than or equal to 1.
  • the preset domain similarity threshold is usually a preset value, for example, it can be 0.8.
  • the initial domain information may be preset initial information.
  • the initial domain information is usually a vector.
  • the above execution subject can calculate the similarity between the vector corresponding to the training sentence in the target domain and the vector corresponding to the initial domain information, such as cosine similarity, the reciprocal of the Euclidean distance, etc., so as to obtain the difference between the training sentence in the target domain and the initial domain information. domain similarity.
  • Step 202 in response to the domain similarity being less than the preset domain similarity threshold, adjust the initial domain information, use the adjusted initial domain information as the initial domain information, and continue to perform the domain calculation step.
  • the initial domain information is adjusted by a preset adjustment method, and the domain calculation step is continued on the adjusted initial domain information.
  • the preset adjustment method may be to adjust the vector corresponding to the initial domain information by using a gradient descent method.
  • the target domain training sentence since the target domain training sentence usually includes the information of the task indicated by the candidate domain information, the target domain training sentence is used to train the initial domain information to obtain the candidate domain information, which can make the candidate domain information more accurate. Indicate the task. Therefore, the necessary information corresponding to the task can be accurately obtained, which helps to further improve the efficiency of the dialogue.
  • the candidate slot information in the candidate slot information set is obtained through the following steps: a slot calculation step: determining the similarity of the slot between the target slot training statement and the initial slot information
  • the initial slot information is determined as candidate slot information in response to the slot similarity being greater than or equal to the preset slot similarity threshold.
  • the target slot training sentence includes the slot information described by the candidate slot information.
  • the initial slot information is adjusted, the adjusted initial slot information is used as the initial slot information, and the slot calculation step is continued.
  • the manner of obtaining the candidate slot information in the candidate slot information set is basically similar to the foregoing manner of obtaining the candidate field information in the candidate field information set, and details are not described here.
  • FIG. 3 is an implementation flowchart of a method for tracking a dialog state provided by another embodiment of the present application.
  • the dialog state tracking method provided in this embodiment is a further refinement of step 103 .
  • step 103 may include steps 301 and 302 . Details are as follows:
  • Step 301 input current field information and historical field state information into a pre-trained first neural network model, obtain current field state information, and input current slot information and historical slot state information into a pre-trained second neural network model, Get the current slot status information.
  • the first neural network model is usually used to represent the correspondence between current domain information, historical domain state information and domain state information.
  • the first neural network model may be a correspondence table that is generated based on statistics of a large number of states of a large number of conversations, and stores a plurality of correspondence between current domain information, historical domain state information and domain state information, or may be The model obtained after training the initial model (eg, convolutional neural network, residual network, etc.) using machine learning methods based on the training samples.
  • the second neural network model is usually used to represent the correspondence between current slot information, historical slot state information and slot state information.
  • the second neural network model may be a correspondence table that is generated based on statistics on a large number of states of a large number of conversations, and stores a plurality of current slot information, historical slot state information, and correspondence between the slot state information and the corresponding relationship table, It can also be a model obtained after training an initial model (for example, a convolutional neural network, a residual network, etc.) by using a machine learning method based on the training samples.
  • the first neural network model may be a long short-term memory network model (Long Short-Term Memory, LSTM) model
  • the second neural network model may also be an LSTM model.
  • Step 302 combine the current domain state information and the current slot state information to generate the dialog state information of the current round of dialog.
  • the current domain state information and the current slot state information can be directly spliced into the dialog state information.
  • the current domain status information is: buy a ticket.
  • the current slot status information is: price, cheap.
  • the current domain status information and the current slot status information can be spliced to obtain: buy air tickets - price - cheap.
  • the current domain state information and the current slot state information are usually spliced into a vector, and the dialog state information is usually obtained by directly splicing the two vectors. For example, if the current domain state information is [1, 3, 5] and the current slot state information is [2, 4, 6], after the two vectors are spliced, the dialog state information can be obtained as [1, 3, 5, 2, 4, 6].
  • two neural network models are used to determine the domain state information and slot state information of the current conversation respectively, so that the domain state information and slot state information of the current conversation can be captured more accurately, which is helpful to achieve more accurate Dialogue state tracking can be carried out to further improve the efficiency of human-machine dialogue.
  • FIG. 4 is a schematic diagram of a dialog state tracking method provided by another embodiment of the present application. Details are as follows:
  • the candidate domain information set select the candidate domain information that matches the user's current input and the robot's last round of replies to obtain the current domain information. And, from the candidate slot information set, the candidate slot information that matches the current input of the user and the robot's last round reply is selected to obtain the current slot information.
  • the current input of the user is the current input sentence input by the user in the current round of dialog.
  • the last turn of the robot is the system response sentence of the terminal device for the last input sentence.
  • the LSTM model used to obtain the current domain state information is the first neural network model
  • the LSTM model used to obtain the current slot state information is the second neural network model
  • the current domain state information and the current slot state information can be directly spliced into the dialog state information.
  • the current domain status information is: buy a ticket.
  • the current slot status information is: price, cheap.
  • the current domain status information and the current slot status information can be spliced to obtain: buy air tickets - price - cheap.
  • the current domain state information and the current slot state information are usually spliced into a vector, and the dialog state information is usually obtained by directly splicing the two vectors. For example, if the current domain state information is [1, 3, 5] and the current slot state information is [2, 4, 6], after the two vectors are spliced, the dialog state information can be obtained as [1, 3, 5, 2, 4, 6].
  • the terminal device in the man-machine dialogue, can obtain a dialogue state information for each round of dialogue between the user and the terminal device.
  • the terminal device can upload the dialogue status information of each round of dialogue to the blockchain to ensure its security and fairness and transparency to users.
  • the user equipment can download the conversation state information from the blockchain, so as to verify whether the conversation state information has been tampered with.
  • the blockchain referred to in this example is a new application mode of computer technology such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm.
  • Blockchain essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information to verify its Validity of information (anti-counterfeiting) and generation of the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.
  • FIG. 5 is a structural block diagram of an apparatus 500 for determining a dialog state provided by an embodiment of the present application.
  • each unit included in the terminal device is used to execute each step in the embodiment corresponding to FIG. 1 to FIG. 4 .
  • FIG. 1 to FIG. 4 and the related descriptions in the embodiments corresponding to FIG. 1 to FIG. 4 .
  • the dialog state determining apparatus 500 includes:
  • the sentence receiving unit 501 is configured to, in response to receiving the current input sentence input by the user in the current round of dialogue, obtain a system response sentence for the previous input sentence, wherein the previous input sentence is the one input by the user in the previous round of dialogue statement;
  • the information selection unit 502 is configured to select, from the pre-built candidate domain information set, the candidate domain information that matches the current input sentence and the system response sentence as the current domain information, and from the pre-built candidate slot information set, select the candidate domain information that matches the current input sentence and the system response sentence.
  • the candidate slot information matching the input statement and the system response statement is used as the current slot information;
  • the information determination unit 503 is used to input the current domain information, historical domain status information, current slot information and historical slot status information into the pre-trained dialogue state model to obtain the dialogue state information of the current round of dialogue, wherein the dialogue state information Including current domain status information and current slot status information.
  • the information determination unit 503 is specifically configured to:
  • the information selection unit 502 selects, from the pre-built candidate domain information set, the candidate domain information that matches the current input sentence and the system response sentence, as the current domain information, including: from the candidate domain information set, Select the candidate domain information that is most similar to the semantics of the combined statement as the current domain information, wherein the combined statement is a statement generated by combining the current input statement and the system response statement; and
  • the candidate slot information that matches the current input statement and the system response statement, as the current slot information, including:
  • the candidate slot information that is most similar to the semantics of the combined sentence is selected as the current slot information.
  • the candidate domain information that is most similar to the semantics of the combined sentence is selected as the current domain information, including:
  • For the candidate field information in the candidate field information set determine the semantic similarity between the vector corresponding to the candidate field information and the vector corresponding to the combined sentence;
  • the candidate domain information with the largest corresponding semantic similarity is selected as the current domain information.
  • the candidate slot information that is most similar to the semantics of the combined statement is selected as the current slot information, including:
  • For the candidate slot information in the candidate slot information set determine the semantic similarity between the vector corresponding to the candidate slot information and the vector corresponding to the combined sentence;
  • the candidate slot information with the largest corresponding semantic similarity is selected as the current slot information.
  • the candidate domain information in the candidate domain information set is obtained through the following steps:
  • Domain calculation step determine the domain similarity between the target domain training sentence and the initial domain information, and in response to the domain similarity being greater than or equal to a preset domain similarity threshold, determine the initial domain information as candidate domain information, wherein the target domain training The statement includes information about the task indicated by the candidate domain information;
  • the initial domain information is adjusted, and the adjusted initial domain information is used as the initial domain information, and the domain calculation step is continued.
  • the candidate slot information in the candidate slot information set is obtained through the following steps:
  • Slot calculation step determine the slot similarity between the target slot training sentence and the initial slot information, and determine the initial slot information as a candidate slot in response to the slot similarity being greater than or equal to a preset slot similarity threshold bit information, wherein the target slot training statement includes the slot information described by the candidate slot information;
  • the initial slot information is adjusted, the adjusted initial slot information is used as the initial slot information, and the slot calculation step is continued.
  • the dialogue state model is obtained by training the following steps:
  • the training samples in the training sample set include domain information, historical domain status information, slot information, historical slot status information and corresponding dialogue status information;
  • the present application can save computing resources and improve the efficiency of obtaining necessary information of tasks to be performed. It helps to improve the efficiency of human-machine dialogue.
  • each unit is used to execute each step in the embodiment corresponding to FIG. 1 to FIG. 4 , and for the embodiment corresponding to FIG. 1 to FIG. 4
  • Each step of the above has been explained in detail in the above-mentioned embodiments.
  • FIG. 6 is a structural block diagram of a terminal device provided by another embodiment of the present application.
  • the terminal device 600 of this embodiment includes: a processor 601 , a memory 602 , and a computer program 603 stored in the memory 602 and executable on the processor 601 , such as a program of a dialog state determination method.
  • the processor 601 executes the computer program 603, it implements the steps in the various embodiments of the above-mentioned dialog state determination methods, for example, steps 101 to 103 shown in FIG. 1, or 201 to 202 shown in FIG. 301 to 302.
  • the processor 601 executes the computer program 603, the functions of the units in the embodiment corresponding to FIG. 5 are implemented, for example, the functions of the units 501 to 503 shown in FIG. description, which is not repeated here.
  • the computer program 603 may be divided into one or more units, and the one or more units are stored in the memory 602 and executed by the processor 601 to complete the present application.
  • One or more units may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program 603 in the terminal device 600 .
  • the computer program 603 can be divided into a sentence receiving unit, an information selecting unit, and an information determining unit, and the specific functions of each unit are as above.
  • the terminal device may include, but is not limited to, the processor 601 and the memory 602 .
  • FIG. 6 is only an example of the terminal device 600, and does not constitute a limitation on the terminal device 600, and may include more or less components than the one shown, or combine some components, or different components
  • the terminal device may also include an input and output device, a network access device, a bus, and the like.
  • the so-called processor 601 may be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application-specific integrated circuits (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 602 may be an internal storage unit of the terminal device 600 , such as a hard disk or a memory of the terminal device 600 .
  • the memory 602 may also be an external storage device of the terminal device 600, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital) device equipped on the terminal device 600 Digital, SD) card, flash memory card (Flash Card), etc.
  • the memory 602 may also include both an internal storage unit of the terminal device 600 and an external storage device.
  • the memory 602 is used to store computer programs and other programs and data required by the terminal device.
  • the memory 602 may also be used to temporarily store data that has been or will be output.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit.
  • the above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.
  • the integrated modules if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer-readable storage medium.
  • the computer-readable storage medium may be non-volatile or volatile.
  • this application can implement all or part of the processes in the methods of the above embodiments, and it can also be completed by instructing the relevant hardware through a computer program.
  • the computer program can be stored in a computer-readable storage medium, and the computer program can be When executed by the processor, the steps of the foregoing method embodiments may be implemented.
  • the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file or some intermediate forms, and the like.
  • Computer-readable storage media may include: any entity or device capable of carrying computer program codes, recording media, USB flash drives, removable hard disks, magnetic disks, optical discs, computer memory, read-only memory (ROM, Read-Only Memory) Memory), random access memory (RAM, Random Access Memory), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased in accordance with the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to the legislation and patent practice, the computer-readable storage medium Electric carrier signals and telecommunication signals are not included.

Abstract

一种对话状态确定方法、终端设备及存储介质，适用于人工智能技术领域，该方法包括：响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句（101）；从预先构建的候选领域信息集中，选取与当前输入语句和系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与当前输入语句和系统响应语句匹配的候选槽位信息，作为当前槽位信息（102）；将当前领域信息、历史领域状态信息、当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息（103）。

Description

对话状态确定方法、终端设备及存储介质
本申请要求于2020年12月28日提交中国专利局、申请号为202011586720.2，发明名称为“对话状态确定方法、终端设备及存储介质”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
本申请涉及人工智能技术领域，尤其涉及一种对话状态确定方法、终端设备及存储介质。
 背景技术
人机交互中，用户与智能设备的一问一答通常被称之为一轮对话。目前，智能设备通常可以通过多轮对话，获取必要信息以最终得到明确的用户指令。实践中，多轮对话通常与一件任务的执行相对应。如，可以通过多轮对话，获取用于执行购买机票任务的必要信息，从而基于所获取的必要信息得到明确的用户指令，以实现准确执行购买机票的任务。
相关技术中，需要通过多轮对话获取所需执行的任务的必要信息。
 技术问题
综上，发明人意识到，需要通过多轮对话获取所需执行的任务的必要信息，而相关技术中，获取所需执行的任务的必要信息的效率不够高。
 技术解决方案
本申请旨在提供一种对话状态确定方法、终端设备及存储介质，其能够在一定程度上提高获取所需执行的任务的必要信息的效率。
根据本申请实施例的第一方面，提供了一种对话状态确定方法，包括：
响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，上一输入语句为用户在上一轮对话中输入的语句；
从预先构建的候选领域信息集中，选取与当前输入语句和系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与当前输入语句和系统响应语句匹配的候选槽位信息，作为当前槽位信息；
将当前领域信息、历史领域状态信息、当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，对话状态信息包括当前领域状态信息和当前槽位状态信息。
根据本申请实施例的第二方面，提供了一种对话状态确定装置，包括：
语句接收单元，用于响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，上一输入语句为用户在上一轮对话中输入的语句；
信息选取单元，用于从预先构建的候选领域信息集中，选取与当前输入语句和系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与当前输入语句和系统响应语句匹配的候选槽位信息，作为当前槽位信息；
信息确定单元，用于将当前领域信息、历史领域状态信息、当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，对话状态信息包括当前领域状态信息和当前槽位状态信息。
根据本申请实施例的第三方面，提供了一种终端设备，包括存储器、处理器以及存储在存储器中并可在终端设备上运行的计算机程序，处理器执行计算机程序时实现如下步骤：
响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，所述上一输入语句为用户在上一轮对话中输入的语句；
从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息；
将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，所述对话状态信息包括当前领域状态信息和当前槽位状态信息。
根据本申请实施例的第四方面，提供了一种计算机可读存储介质，其存储有计算机程序，当计算机程序被处理器执行时，实现如下步骤：
响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，所述上一输入语句为用户在上一轮对话中输入的语句；
从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息；
将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，所述对话状态信息包括当前领域状态信息和当前槽位状态信息。
 有益效果
本申请实施例与相关技术相比存在的有益效果是：通过引入领域因素，使得不同领域的信息可以被区分，实现在一次多轮对话中，收集到不同领域任务对应的必要信息。与相关技术中，通过多次多轮对话以获取各领域任务分别对应的必要信息相比，本申请可以节约计算资源，提高获取所需执行的任务的必要信息的效率。有助于提高人机对话效率。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性的，并不能限制本申请。
 附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例或相关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性的前提下，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状态确定方法的实现流程图；
图2是本申请实施例提供的候选领域信息集的构建方法的实现流程图；
图3是本申请另一实施例提供的一种对话状态确定方法的实现流程图；
图4是本申请再一实施例提供的一种对话状态确定方法的实现流程图；
图5是本申请实施例提供的一种对话状态确定装置的结构框图；
图6是本申请实施例提供的一种终端设备的结构框图。
 本发明的实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
本申请实施例所涉及的对话状态确定方法，可以由控制设备或终端（以下称“终端设备”）执行。
请参阅图1，图1示出了本申请实施例提供的一种对话状态确定方法的实现流程图，包括：
步骤101，响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句。
其中，当前输入语句，通常为用户在当前轮对话中输入的语句。上一输入语句，通常为用户在上一轮对话中输入的语句。系统响应语句，通常为终端设备针对用户在上一轮对话中输入的语句所回应的语句。这里，用户输入某一语句后，系统针对该语句做出回应，一输入一回应为一轮对话。
这里，上述执行主体可以接收到用户输入的当前输入语句，以及在接收到当前输入语句时，调出上一轮对话中，执行主体针对用户输入的上一输入语句所回应的系统响应语句。
实践中，上述当前输入语句和上述系统响应语句通常可以是语音形式或文字形式。作为示例，若当前输入语句为语音形式，则执行主体可以先将语音转换成文字，然后将文字转换成便于数据处理的向量。若当前输入语句为文字形式，则执行主体可以直接将文字转换成便于数据处理的向量。实际应用中，执行主体可以采用双向长短期记忆网络（Bi-directional Long Short-Term Memory，Bi-LSTM）将语句对应文字转换成向量。执行主体还可以采用word2vec模型将语句中的词转换成向量，然后将各个词对应的向量拼接起来，形成维度为所有词向量的维度之和的向量，得到语句对应的向量。其中，word2vec模型用于表征语句中的词与词向量的对应关系，word2vec模型是本领域技术人员常用的用于将词转化成词向量的神经网络模型，这里不做赘述。
步骤102，从预先构建的候选领域信息集中，选取与当前输入语句和系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与当前输入语句和系统响应语句匹配的候选槽位信息，作为当前槽位信息。
其中，领域信息通常是用于指示目标任务的信息。目标任务可以是预先设定的各种任务。作为示例，任务可以是贷款、还款、买机票等。候选领域信息通常是用于候选的领域信息。
其中，槽位信息通常包括槽位和槽位值。这里，槽位通常是指需要向用户收集的关键信息。槽位通常具有槽位值，且槽位值通常是用户表达的关键信息的具体取值。作为一个示例，槽位可以是价格，槽位值可以是价格的取值。作为另一个示例，槽位可以是出发地，槽位值可以是上海。实践中，填槽通常是收集用户信息的过程，用于将模糊或缺失的用户意图补全。
这里，上述执行主体可以通过如下方式得到当前领域信息和当前槽位信息：从候选领域信息集中，选取与当前输入语句和系统响应语句具有最多重复字的候选领域信息，作为当前领域信息。以及从候选槽位信息集中，选取与当前输入语句和系统响应语句具有最多重复字的候选槽位信息，作为当前槽位信息。
步骤103，将当前领域信息、历史领域状态信息、当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息。
其中，对话状态信息包括当前领域状态信息和当前槽位状态信息。当前领域状态信息通常为当前轮对话以及当前轮对话之前的各轮对话综合对应的领域信息。当前领域状态信息中可以包括多个用于指示目标任务的信息，即，当前领域状态信息可以同时指示多个目标任务。当前槽位状态信息通常为当前轮对话以及当前轮对话之前的各轮对话综合对应的槽位信息。历史领域状态信息，通常为在当前轮对话之前的各轮对话综合对应的领域信息。历史槽位状态信息，通常为在当前轮对话之前的各轮对话综合对应的槽位信息。
其中，对话状态模型通常用于表征当前领域信息、历史领域状态信息、当前槽位信息和历史槽位状态信息与当前对话状态的对应关系。
具体的，对话状态模型可以是基于对大量对话的大量状态进行统计而生成的、存储有多个当前领域信息、历史领域状态信息、当前槽位信息和历史槽位状态信息与对话状态信息的对应关系的对应关系表，也可以是基于训练样本，利用机器学习方法，对初始模型（例如卷积神经网络（Convolutional Neural Network，CNN）、残差网络（ResNet）等）进行训练后得到的模型。
本实施例中，通过引入领域因素，使得不同领域的信息可以被区分，实现在一次多轮对话中，收集到不同领域任务对应的必要信息。与相关技术中，通过多次多轮对话以获取各领域任务分别对应的必要信息相比，本申请可以节约计算资源，提高获取所需执行的任务的必要信息的效率。有助于提高人机对话效率。需要指出的是，对话状态信息是基于当前轮对话以及当前轮对话之前的各轮对话综合分析得到，针对每一轮对话，可以得到对应的对话状态信息，通过一次多轮对话可以得到指示多个任务的对话状态信息，可以提高获取所需执行的任务的必要信息的效率。有助于提高人机对话效率。
在本实施例的一些可选的实现方式中，上述执行主体也可以通过如下方式得到当前领域信息和当前槽位信息：首先，从候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息。然后，从候选槽位信息集中，选取与组合语句的语义最相近的候选槽位信息，作为当前槽位信息。
其中，组合语句可以为将当前输入语句和系统响应语句组合生成的语句。举例来说，若当前输入语句为：帮忙买张去深圳的机票，针对上一轮输入的系统响应语句为：需要我帮忙干点什么呢。则组合语句可以为：需要我帮忙干点什么呢，帮忙买张去深圳的机票。
这里，上述执行主体可以采用语义相似度算法，如，深度语义模型算法(Deep Structured Sematic models，DSSM)、卷积潜在语义模型(Convolutional Latent Semantic Model，CLSM)等，计算组合语句与各候选领域信息之间的相似度，然后，选取对应相似度最大的候选领域信息作为当前领域信息。以及采用上述语义相似度算法，计算组合语句与各候选槽位信息之间的相似度，然后，选取对应相似度最大的候选槽位信息作为当前槽位信息。
本实现方式中，选取语义最相近的候选领域信息作为当前领域信息，且选取语义最相近的候选槽位信息作为当前槽位信息，可以更准确地捕捉到用户真实意图，从而进一步提高人机对话的效率。
在本实施例的一些可选的实现方式中，上述从候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，包括：首先，针对候选领域信息集中的候选领域信息，确定该候选领域信息对应的向量与组合语句对应的向量之间的语义相似度。然后，从候选领域信息集中，选取对应语义相似度最大的候选领域信息，作为当前领域信息。
其中，语义相似度通常是用于描述数据之间的语义相关性的度量值。语义相似度的取值通常大于0且小于或等于1。其中，语义相似度可以包括以下任意一项或多项：余弦相似度，欧氏距离的倒数等。
这里，上述执行主体可以计算各候选领域信息的向量与组合语句的向量之间的语义相似度，然后选取对应语义相似度最大的候选领域信息作为当前领域信息。
本实现方式中，直接将两个向量之间的相似度，作为候选领域信息与组合语句之间的语义相似度，计算复杂度低，可以在保障获取所需执行的任务的必要信息的效率的同时，节约计算资源。
在本实施例的一些可选的实现方式中，从候选槽位信息集中，选取与组合语句的语义最相近的候选槽位信息，作为当前槽位信息，包括：针对候选槽位信息集中的候选槽位信息，确定该候选槽位信息对应的向量与组合语句对应的向量之间的语义相似度。从候选槽位信息集中，选取对应语义相似度最大的候选槽位信息，作为当前槽位信息。
这里，上述执行主体可以计算各候选槽位信息的向量与组合语句的向量之间的语义相似度，然后选取对应语义相似度最大的候选槽位信息作为当前槽位信息。
本实现方式中，直接将两个向量之间的相似度，作为候选槽位信息与组合语句之间的语义相似度，计算复杂度低，可以在保障获取所需执行的任务的必要信息的效率的同时，进一步节约计算资源。
在本实施例的一些可选的实现方式中，对话状态模型通过如下步骤训练得到：首先，获取训练样本集，训练样本集中的训练样本包括领域信息、历史领域状态信息、槽位信息、历史槽位状态信息和对应的对话状态信息。然后，将训练样本集中的训练样本的领域信息、历史领域状态信息、槽位信息、历史槽位状态信息作为输入，将与输入的领域信息、历史领域状态信息、槽位信息、历史槽位状态信息对应的对话状态信息作为期望输出，训练得到对话状态模型。
其中，对话状态信息包括当前领域状态信息和当前槽位状态信息。当前领域状态信息通常为当前轮对话以及当前轮对话之前的各轮对话综合对应的领域信息。当前领域状态信息中可以包括多个用于指示目标任务的信息，即，当前领域状态信息可以同时指示多个目标任务。当前槽位状态信息通常为当前轮对话以及当前轮对话之前的各轮对话综合对应的槽位信息。历史领域状态信息，通常为在当前轮对话之前的各轮对话综合对应的领域信息。历史槽位状态信息，通常为在当前轮对话之前的各轮对话综合对应的槽位信息。
本实现方式中，训练得到的对话状态模型，可以针对所输入的当前领域信息、历史领域状态信息、当前槽位信息和历史槽位状态信息，输出当前轮对话的当前领域状态信息和当前槽位状态信息。只需训练一个模型即可实现同时获取到两种信息，即，当前领域状态信息和当前槽位状态信息。与相关技术中，需要训练至少两个模型以分别获取到当前领域状态信息和当前槽位状态信息相比，本申请能够加快模型训练效率。
请参阅图2，图2是本申请另一实施例提供的候选领域信息集的构建方法的实现流程图。详述如下：
步骤201，领域计算步骤：确定目标领域训练语句与初始领域信息之间的领域相似度，响应于领域相似度大于或等于预设领域相似度阈值，将初始领域信息确定为候选领域信息。
其中，目标领域训练语句包括该候选领域信息所指示的任务的信息。作为示例，目标领域训练语句可以为：帮忙买去深圳的机票。候选领域信息所指示的任务为：买机票。
其中，领域相似度通常是用于描述两个信息所指向的任务的相似程度的数值。领域相似度的取值通常大于0且小于等于1。预设领域相似度阈值通常是预先设定的数值，如，可以为0.8。
其中，初始领域信息可以是预先设定的初始信息，实践中，初始领域信息通常为一向量。上述执行主体可以计算目标领域训练语句对应的向量与初始领域信息对应的向量之间的相似度，如，余弦相似度、欧氏距离的倒数等，实现得到目标领域训练语句与初始领域信息之间的领域相似度。
步骤202，响应于领域相似度小于预设领域相似度阈值，调整初始领域信息，将调整后的初始领域信息作为初始领域信息，继续执行领域计算步骤。
这里，在领域相似度小于预设领域相似度阈值时，采用预设调整方式调整初始领域信息，以及对调整后的初始领域信息继续执行领域计算步骤。作为示例，预设调整方式可以是采用梯度下降方式调整初始领域信息对应的向量。
本实施例中，由于目标领域训练语句通常包括该候选领域信息所指示的任务的信息，采用目标领域训练语句对初始领域信息进行训练，以得到候选领域信息，可以使得候选领域信息能够更加准确地指示任务。从而准确地获取到任务对应的必要信息，有助于进一步提高对话效率。
在本实施例的一些可选的实现方式中，候选槽位信息集中的候选槽位信息通过如下步骤得到：槽位计算步骤：确定目标槽位训练语句与初始槽位信息之间的槽位相似度，响应于槽位相似度大于或等于预设槽位相似度阈值，将初始槽位信息确定为候选槽位信息。目标槽位训练语句包括该候选槽位信息所描述的槽位信息。响应于槽位相似度小于预设槽位相似度阈值，调整初始槽位信息，将调整后的初始槽位信息作为初始槽位信息，继续执行槽位计算步骤。
本实现方式中，得到候选槽位信息集中的候选槽位信息的方式，与前述获取候选领域信息集中的候选领域信息的方式基本类似，这里不做赘述。
请参阅图3，图3是本申请另一实施例提供的对话状态追踪方法的实现流程图。相对于图1对应的实施例，本实施例提供的对话状态追踪方法是对步骤103的进一步细化。这里，步骤103可以包括步骤301和302。详述如下：
步骤301，将当前领域信息和历史领域状态信息输入预先训练的第一神经网络模型，得到当前领域状态信息，以及将当前槽位信息和历史槽位状态信息输入预先训练的第二神经网络模型，得到当前槽位状态信息。
其中，第一神经网络模型通常用于表征当前领域信息、历史领域状态信息与领域状态信息的对应关系。具体的，第一神经网络模型可以是基于对大量对话的大量状态进行统计而生成的、存储有多个当前领域信息、历史领域状态信息与领域状态信息的对应关系的对应关系表，也可以是基于训练样本，利用机器学习方法，对初始模型（例如，卷积神经网络、残差网络等）进行训练后得到的模型。
其中，第二神经网络模型通常用于表征当前槽位信息、历史槽位状态信息与槽位状态信息的对应关系。具体的，第二神经网络模型可以是基于对大量对话的大量状态进行统计而生成的、存储有多个当前槽位信息、历史槽位状态信息与槽位状态信息的对应关系的对应关系表，也可以是基于训练样本，利用机器学习方法，对初始模型（例如，卷积神经网络、残差网络等）进行训练后得到的模型。
实践中，第一神经网络模型可以为长短期记忆网络模型（Long Short-Term Memory，LSTM）模型，第二神经网络模型也可以为LSTM模型。
步骤302，将当前领域状态信息和当前槽位状态信息，组合生成当前轮对话的对话状态信息。
这里，可以直接将当前领域状态信息和当前槽位状态信息拼接成对话状态信息。作为示例，若当前领域状态信息为：买机票。当前槽位状态信息为：价格，便宜。则，可以将当前领域状态信息和当前槽位状态信息进行拼接，得到：买机票-价格-便宜。具体实现时，当前领域状态信息和当前槽位状态信息拼接通常为向量，且通常是通过将两个向量直接拼接，得到对话状态信息。如，若当前领域状态信息为[1，3，5]，当前槽位状态信息为[2，4，6]，则两个向量拼接后，可以得到对话状态信息为[1，3，5，2，4，6]。
本实施例中，分两个神经网络模型分别确定当前对话的领域状态信息和槽位状态信息，可以实现更准确地捕捉到当前对话的领域状态信息和槽位状态信息，有助于实现更准确地进行对话状态追踪，从而进一步提高人机对话效率。
请参阅图4，图4是本申请另一实施例提供的对话状态追踪方法的示意图。详述如下：
首先，从候选领域信息集中，选取出与用户当前输入和机器人上轮回复匹配的候选领域信息，得到当前领域信息。以及，从候选槽位信息集中，选取出与用户当前输入和机器人上轮回复匹配的候选槽位信息，得到当前槽位信息。
这里，用户当前输入为用户在当前轮对话中输入的当前输入语句。机器人上轮回复为终端设备针对上一输入语句的系统响应语句。
然后，将当前领域信息和历史领域状态信息输入LSTM模型，得到当前领域状态信息，以及将当前槽位信息和历史槽位状态信息输入另一LSTM模型，得到当前槽位状态信息。
这里，用于得到当前领域状态信息的LSTM模型为第一神经网络模型，用于得到当前槽位状态信息的LSTM模型为第二神经网络模型。
最后，将当前领域状态信息和当前槽位状态信息拼接，得到对话状态信息。
这里，可以直接将当前领域状态信息和当前槽位状态信息拼接成对话状态信息。作为示例，若当前领域状态信息为：买机票。当前槽位状态信息为：价格，便宜。则，可以将当前领域状态信息和当前槽位状态信息进行拼接，得到：买机票-价格-便宜。具体实现时，当前领域状态信息和当前槽位状态信息拼接通常为向量，且通常是通过将两个向量直接拼接，得到对话状态信息。如，若当前领域状态信息为[1，3，5]，当前槽位状态信息为[2，4，6]，则两个向量拼接后，可以得到对话状态信息为[1，3，5，2，4，6]。
在本申请的所有实施例中，人机对话中，终端设备针对用户与终端设备的每一轮对话，可以得到一个对话状态信息。终端设备可以将各轮对话的对话状态信息上传至区块链可保证其安全性和对用户的公正透明性。用户设备可以从区块链中下载得到该对话状态信息，以便查证对话状态信息是否被篡改。本示例所指区块链是分布式数据存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链(Blockchain)，本质上是一个去中心化的数据库，是一串使用密码学方法相关联产生的数据块，每一个数据块中包含了一批次网络交易的信息，用于验证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。
请参阅图5，图5是本申请实施例提供的一种对话状态确定装置500的结构框图。本实施例中该终端设备包括的各单元用于执行图1至图4对应的实施例中的各步骤。具体请参阅图1至图4以及图1至图4所对应的实施例中的相关描述。为了便于说明，仅示出了与本实施例相关的部分。参见图5，对话状态确定装置500包括：
语句接收单元501，用于响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，上一输入语句为用户在上一轮对话中输入的语句；
信息选取单元502，用于从预先构建的候选领域信息集中，选取与当前输入语句和系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与当前输入语句和系统响应语句匹配的候选槽位信息，作为当前槽位信息；
信息确定单元503，用于将当前领域信息、历史领域状态信息、当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，对话状态信息包括当前领域状态信息和当前槽位状态信息。
作为本申请一实施例，信息确定单元503具体用于：
将当前领域信息和历史领域状态信息输入预先训练的第一神经网络模型，得到当前领域状态信息，以及将当前槽位信息和历史槽位状态信息输入预先训练的第二神经网络模型，得到当前槽位状态信息；
将当前领域状态信息和当前槽位状态信息，组合生成当前轮对话的对话状态信息。
作为本申请一实施例，信息选取单元502中，从预先构建的候选领域信息集中，选取与当前输入语句和系统响应语句匹配的候选领域信息，作为当前领域信息，包括：从候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，其中，组合语句为将当前输入语句和系统响应语句组合生成的语句；以及
信息选取单元502中，从预先构建的候选槽位信息集中，选取与当前输入语句和系统响应语句匹配的候选槽位信息，作为当前槽位信息，包括：
从候选槽位信息集中，选取与组合语句的语义最相近的候选槽位信息，作为当前槽位信息。
作为本申请一实施例，从候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，包括：
针对候选领域信息集中的候选领域信息，确定该候选领域信息对应的向量与组合语句对应的向量之间的语义相似度；
从候选领域信息集中，选取对应语义相似度最大的候选领域信息，作为当前领域信息。
作为本申请一实施例，从候选槽位信息集中，选取与组合语句的语义最相近的候选槽位信息，作为当前槽位信息，包括：
针对候选槽位信息集中的候选槽位信息，确定该候选槽位信息对应的向量与组合语句对应的向量之间的语义相似度；
从候选槽位信息集中，选取对应语义相似度最大的候选槽位信息，作为当前槽位信息。
作为本申请一实施例，候选领域信息集中的候选领域信息通过如下步骤得到：
领域计算步骤：确定目标领域训练语句与初始领域信息之间的领域相似度，响应于领域相似度大于或等于预设领域相似度阈值，将初始领域信息确定为候选领域信息，其中，目标领域训练语句包括该候选领域信息所指示的任务的信息；
响应于领域相似度小于预设领域相似度阈值，调整初始领域信息，将调整后的初始领域信息作为初始领域信息，继续执行领域计算步骤。
作为本申请一实施例，候选槽位信息集中的候选槽位信息通过如下步骤得到：
槽位计算步骤：确定目标槽位训练语句与初始槽位信息之间的槽位相似度，响应于槽位相似度大于或等于预设槽位相似度阈值，将初始槽位信息确定为候选槽位信息，其中，目标槽位训练语句包括该候选槽位信息所描述的槽位信息；
响应于槽位相似度小于预设槽位相似度阈值，调整初始槽位信息，将调整后的初始槽位信息作为初始槽位信息，继续执行槽位计算步骤。
作为本申请一实施例，对话状态模型通过如下步骤训练得到：
获取训练样本集，训练样本集中的训练样本包括领域信息、历史领域状态信息、槽位信息、历史槽位状态信息和对应的对话状态信息；
将训练样本集中的训练样本的领域信息、历史领域状态信息、槽位信息、历史槽位状态信息作为输入，将与输入的领域信息、历史领域状态信息、槽位信息、历史槽位状态信息对应的对话状态信息作为期望输出，训练得到对话状态模型。
本实施例提供的装置，通过引入领域因素，使得不同领域的信息可以被区分，实现在一次多轮对话中，收集到不同领域任务对应的必要信息。与相关技术中，通过多次多轮对话以获取各领域任务分别对应的必要信息相比，本申请可以节约计算资源，提高获取所需执行的任务的必要信息的效率。有助于提高人机对话效率。
应当理解的是，图5示出的对话状态确定装置的结构框图中，各单元用于执行图1至图4对应的实施例中的各步骤，而对于图1至图4对应的实施例中的各步骤已在上述实施例中进行详细解释，具体请参阅图1至图4以及图1至图4所对应的实施例中的相关描述，此处不再赘述。
图6是本申请另一实施例提供的一种终端设备的结构框图。如图6所示，该实施例的终端设备600包括：处理器601、存储器602以及存储在存储器602中并可在处理器601上运行的计算机程序603，例如对话状态确定方法的程序。处理器601执行计算机程序603时实现上述各个对话状态确定方法各实施例中的步骤，例如图1所示的步骤101至步骤103，或者图2所示的201至202，或者图3所示的301至302。或者，处理器601执行计算机程序603时实现上述图5对应的实施例中各单元的功能，例如，图5所示的单元501至503的功能，具体请参阅图5对应的实施例中的相关描述，此处不赘述。
示例性的，计算机程序603可以被分割成一个或多个单元，一个或者多个单元被存储在存储器602中，并由处理器601执行，以完成本申请。一个或多个单元可以是能够完成特定功能的一系列计算机程序指令段，该指令段用于描述计算机程序603在终端设备600中的执行过程。例如，计算机程序603可以被分割成语句接收单元、信息选取单元、信息确定单元，各单元具体功能如上。
终端设备可包括，但不仅限于，处理器601、存储器602。本领域技术人员可以理解，图6仅仅是终端设备600的示例，并不构成对终端设备600的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如终端设备还可以包括输入输出设备、网络接入设备、总线等。
所称处理器601可以是中央处理单元（Central Processing Unit，CPU），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。
存储器602可以是终端设备600的内部存储单元，例如终端设备600的硬盘或内存。存储器602也可以是终端设备600的外部存储设备，例如终端设备600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，存储器602还可以既包括终端设备600的内部存储单元也包括外部存储设备。存储器602用于存储计算机程序以及终端设备所需的其他程序和数据。存储器602还可以用于暂时地存储已经输出或者将要输出的数据。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
集成的模块如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读存储介质中。其中，计算机可读存储介质可以是非易失性的，也可以是易失性的。基于这样的理解，本申请实现上述实施例方法中的全部或部分流程，也可以通过计算机程序来指令相关的硬件来完成，计算机程序可存储于一计算机可读存储介质中，该计算机程序在被处理器执行时，可实现上述各个方法实施例的步骤。其中，计算机程序包括计算机程序代码，计算机程序代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。计算机可读存储介质可以包括：能够携带计算机程序代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器（ROM，Read-Only Memory）、随机存取存储器（RAM，Random Access Memory）、电载波信号、电信信号以及软件分发介质等。需要说明的是，计算机可读存储介质包含的内容可以根据司法管辖区内立法和专利实践的要求进行适当的增减，例如在某些司法管辖区，根据立法和专利实践，计算机可读存储介质不包括电载波信号和电信信号。
以上实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种对话状态确定方法，其中，包括：
    响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，所述上一输入语句为用户在上一轮对话中输入的语句；
    从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息；
    将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，所述对话状态信息包括当前领域状态信息和当前槽位状态信息。
  2. 根据权利要求1所述的对话状态确定方法，其中，所述将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，包括：
    将所述当前领域信息和所述历史领域状态信息输入预先训练的第一神经网络模型，得到所述当前领域状态信息，以及将所述当前槽位信息和所述历史槽位状态信息输入预先训练的第二神经网络模型，得到所述当前槽位状态信息；
    将所述当前领域状态信息和所述当前槽位状态信息，组合生成当前轮对话的对话状态信息。
  3. 根据权利要求1所述的对话状态确定方法，其中，
    所述从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，包括：
    从所述候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，其中，所述组合语句为将所述当前输入语句和所述系统响应语句组合生成的语句；以及
    所述从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息，包括：
    从所述候选槽位信息集中，选取与所述组合语句的语义最相近的候选槽位信息，作为当前槽位信息。
  4. 根据权利要求3所述的对话状态确定方法，其中，所述从所述候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，包括：
    针对所述候选领域信息集中的候选领域信息，确定该候选领域信息对应的向量与所述组合语句对应的向量之间的语义相似度；
    从所述候选领域信息集中，选取对应语义相似度最大的候选领域信息，作为当前领域信息。
  5. 根据权利要求3所述的对话状态确定方法，其中，所述从所述候选槽位信息集中，选取与所述组合语句的语义最相近的候选槽位信息，作为当前槽位信息，包括：
    针对所述候选槽位信息集中的候选槽位信息，确定该候选槽位信息对应的向量与所述组合语句对应的向量之间的语义相似度；
    从所述候选槽位信息集中，选取对应语义相似度最大的候选槽位信息，作为当前槽位信息。
  6. 根据权利要求1所述的对话状态确定方法，其中，所述候选领域信息集中的候选领域信息通过如下步骤得到：
    领域计算步骤：确定目标领域训练语句与初始领域信息之间的领域相似度，响应于所述领域相似度大于或等于预设领域相似度阈值，将初始领域信息确定为候选领域信息，其中，所述目标领域训练语句包括该候选领域信息所指示的任务的信息；
    响应于所述领域相似度小于所述预设领域相似度阈值，调整初始领域信息，将调整后的初始领域信息作为初始领域信息，继续执行所述领域计算步骤。
  7. 根据权利要求1所述的对话状态确定方法，其中，所述候选槽位信息集中的候选槽位信息通过如下步骤得到：
    槽位计算步骤：确定目标槽位训练语句与初始槽位信息之间的槽位相似度，响应于所述槽位相似度大于或等于预设槽位相似度阈值，将初始槽位信息确定为候选槽位信息，其中，所述目标槽位训练语句包括该候选槽位信息所描述的槽位信息；
    响应于所述槽位相似度小于所述预设槽位相似度阈值，调整初始槽位信息，将调整后的初始槽位信息作为初始槽位信息，继续执行所述槽位计算步骤。
  8. 根据权利要求1所述的对话状态确定方法，其中，所述对话状态模型通过如下步骤训练得到：
    获取训练样本集，所述训练样本集中的训练样本包括领域信息、历史领域状态信息、槽位信息、历史槽位状态信息和对应的对话状态信息；
    将所述训练样本集中的训练样本的领域信息、历史领域状态信息、槽位信息、历史槽位状态信息作为输入，将与输入的领域信息、历史领域状态信息、槽位信息、历史槽位状态信息对应的对话状态信息作为期望输出，训练得到所述对话状态模型。
  9. 一种对话状态确定装置，其中，包括：
    语句接收单元，用于响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，所述上一输入语句为用户在上一轮对话中输入的语句；
    信息选取单元，用于从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息；
    信息确定单元，用于将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，所述对话状态信息包括当前领域状态信息和当前槽位状态信息。
  10. 一种终端设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其中，所述处理器执行所述计算机程序时，实现如下步骤：
    响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，所述上一输入语句为用户在上一轮对话中输入的语句；
    从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息；
    将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，所述对话状态信息包括当前领域状态信息和当前槽位状态信息。
  11. 根据权利要求10所述的终端设备，其中，所述将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，包括：
    将所述当前领域信息和所述历史领域状态信息输入预先训练的第一神经网络模型，得到所述当前领域状态信息，以及将所述当前槽位信息和所述历史槽位状态信息输入预先训练的第二神经网络模型，得到所述当前槽位状态信息；
    将所述当前领域状态信息和所述当前槽位状态信息，组合生成当前轮对话的对话状态信息。
  12. 根据权利要求10所述的终端设备，其中，
    所述从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，包括：
    从所述候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，其中，所述组合语句为将所述当前输入语句和所述系统响应语句组合生成的语句；以及
    所述从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息，包括：
    从所述候选槽位信息集中，选取与所述组合语句的语义最相近的候选槽位信息，作为当前槽位信息。
  13. 根据权利要求12所述的终端设备，其中，所述从所述候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，包括：
    针对所述候选领域信息集中的候选领域信息，确定该候选领域信息对应的向量与所述组合语句对应的向量之间的语义相似度；
    从所述候选领域信息集中，选取对应语义相似度最大的候选领域信息，作为当前领域信息。
  14. 根据权利要求12所述的终端设备，其中，所述从所述候选槽位信息集中，选取与所述组合语句的语义最相近的候选槽位信息，作为当前槽位信息，包括：
    针对所述候选槽位信息集中的候选槽位信息，确定该候选槽位信息对应的向量与所述组合语句对应的向量之间的语义相似度；
    从所述候选槽位信息集中，选取对应语义相似度最大的候选槽位信息，作为当前槽位信息。
  15. 根据权利要求10所述的终端设备，其中，所述候选领域信息集中的候选领域信息通过如下步骤得到：
    领域计算步骤：确定目标领域训练语句与初始领域信息之间的领域相似度，响应于所述领域相似度大于或等于预设领域相似度阈值，将初始领域信息确定为候选领域信息，其中，所述目标领域训练语句包括该候选领域信息所指示的任务的信息；
    响应于所述领域相似度小于所述预设领域相似度阈值，调整初始领域信息，将调整后的初始领域信息作为初始领域信息，继续执行所述领域计算步骤。
  16. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其中，所述计算机程序被处理器执行时，实现如下步骤：
    响应于接收到用户在当前轮对话中输入的当前输入语句，获取针对上一输入语句的系统响应语句，其中，所述上一输入语句为用户在上一轮对话中输入的语句；
    从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，以及从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息；
    将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，其中，所述对话状态信息包括当前领域状态信息和当前槽位状态信息。
  17. 根据权利要求16所述的计算机可读存储介质，其中，所述将所述当前领域信息、历史领域状态信息、所述当前槽位信息和历史槽位状态信息，输入预先训练的对话状态模型，得到当前轮对话的对话状态信息，包括：
    将所述当前领域信息和所述历史领域状态信息输入预先训练的第一神经网络模型，得到所述当前领域状态信息，以及将所述当前槽位信息和所述历史槽位状态信息输入预先训练的第二神经网络模型，得到所述当前槽位状态信息；
    将所述当前领域状态信息和所述当前槽位状态信息，组合生成当前轮对话的对话状态信息。
  18. 根据权利要求16所述的计算机可读存储介质，其中，
    所述从预先构建的候选领域信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选领域信息，作为当前领域信息，包括：
    从所述候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，其中，所述组合语句为将所述当前输入语句和所述系统响应语句组合生成的语句；以及
    所述从预先构建的候选槽位信息集中，选取与所述当前输入语句和所述系统响应语句匹配的候选槽位信息，作为当前槽位信息，包括：
    从所述候选槽位信息集中，选取与所述组合语句的语义最相近的候选槽位信息，作为当前槽位信息。
  19. 根据权利要求18所述的计算机可读存储介质，其中，所述从所述候选领域信息集中，选取与组合语句的语义最相近的候选领域信息，作为当前领域信息，包括：
    针对所述候选领域信息集中的候选领域信息，确定该候选领域信息对应的向量与所述组合语句对应的向量之间的语义相似度；
    从所述候选领域信息集中，选取对应语义相似度最大的候选领域信息，作为当前领域信息。
  20. 根据权利要求18所述的计算机可读存储介质，其中，所述从所述候选槽位信息集中，选取与所述组合语句的语义最相近的候选槽位信息，作为当前槽位信息，包括：
    针对所述候选槽位信息集中的候选槽位信息，确定该候选槽位信息对应的向量与所述组合语句对应的向量之间的语义相似度；
    从所述候选槽位信息集中，选取对应语义相似度最大的候选槽位信息，作为当前槽位信息。
PCT/CN2021/091132 2020-12-28 2021-04-29 对话状态确定方法、终端设备及存储介质 WO2022142028A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011586720.2A CN112632254A (zh) 2020-12-28 2020-12-28 对话状态确定方法、终端设备及存储介质
CN202011586720.2 2020-12-28

Publications (1)

Publication Number Publication Date
WO2022142028A1 true WO2022142028A1 (zh) 2022-07-07

Family

ID=75286210

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/091132 WO2022142028A1 (zh) 2020-12-28 2021-04-29 对话状态确定方法、终端设备及存储介质

Country Status (2)

Country Link
CN (1) CN112632254A (zh)
WO (1) WO2022142028A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112632254A (zh) * 2020-12-28 2021-04-09 平安科技（深圳）有限公司 对话状态确定方法、终端设备及存储介质
CN113553413A (zh) * 2021-06-30 2021-10-26 北京百度网讯科技有限公司 对话状态的生成方法、装置、电子设备和存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080201135A1 (en) * 2007-02-20 2008-08-21 Kabushiki Kaisha Toshiba Spoken Dialog System and Method
CN110555095A (zh) * 2018-05-31 2019-12-10 北京京东尚科信息技术有限公司 人机对话方法和装置
CN110858226A (zh) * 2018-08-07 2020-03-03 北京京东尚科信息技术有限公司 对话管理方法和装置
CN111090728A (zh) * 2019-12-13 2020-05-01 车智互联（北京）科技有限公司 一种对话状态跟踪方法、装置及计算设备
CN111125326A (zh) * 2019-12-06 2020-05-08 贝壳技术有限公司 用于实现人机对话的方法、装置、介质以及电子设备
CN112632254A (zh) * 2020-12-28 2021-04-09 平安科技（深圳）有限公司 对话状态确定方法、终端设备及存储介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107369443B (zh) * 2017-06-29 2020-09-25 北京百度网讯科技有限公司 基于人工智能的对话管理方法及装置
CN109063035B (zh) * 2018-07-16 2021-11-09 哈尔滨工业大学 一种面向出行领域的人机多轮对话方法
CN111414195B (zh) * 2019-02-03 2022-03-25 北京邮电大学 通用可配置、兼容多模型、领域可迁移的多轮对话系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080201135A1 (en) * 2007-02-20 2008-08-21 Kabushiki Kaisha Toshiba Spoken Dialog System and Method
CN110555095A (zh) * 2018-05-31 2019-12-10 北京京东尚科信息技术有限公司 人机对话方法和装置
CN110858226A (zh) * 2018-08-07 2020-03-03 北京京东尚科信息技术有限公司 对话管理方法和装置
CN111125326A (zh) * 2019-12-06 2020-05-08 贝壳技术有限公司 用于实现人机对话的方法、装置、介质以及电子设备
CN111090728A (zh) * 2019-12-13 2020-05-01 车智互联（北京）科技有限公司 一种对话状态跟踪方法、装置及计算设备
CN112632254A (zh) * 2020-12-28 2021-04-09 平安科技（深圳）有限公司 对话状态确定方法、终端设备及存储介质

Also Published As

Publication number Publication date
CN112632254A (zh) 2021-04-09

Similar Documents

Publication Publication Date Title
US11886998B2 (en) Attention-based decoder-only sequence transduction neural networks
US11425064B2 (en) Customized message suggestion with user embedding vectors
CN111738016B (zh) 多意图识别方法及相关设备
CN111046667B (zh) 一种语句识别方法、语句识别装置及智能设备
WO2022142028A1 (zh) 对话状态确定方法、终端设备及存储介质
CN112860841A (zh) 一种文本情感分析方法、装置、设备及存储介质
CN112163067A (zh) 一种语句答复方法、语句答复装置及电子设备
CN113298152B (zh) 模型训练方法、装置、终端设备及计算机可读存储介质
CN110457449A (zh) 在线训练模型的方法、装置、设备及存储介质
US11741370B2 (en) Transfer learning based on cross-domain homophily influences
CN113656547B (zh) 文本匹配方法、装置、设备及存储介质
WO2022257452A1 (zh) 表情回复方法、装置、设备及存储介质
JP2020064621A (ja) 敵対的生成ネットワークを用いるユーザフレンドリな説明生成
US11475297B2 (en) Cross-domain homophily quantification for transfer learning
CN111046653B (zh) 一种语句识别方法、语句识别装置及智能设备
CN115237802A (zh) 基于人工智能的模拟测试方法及相关设备
CN112836502A (zh) 一种金融领域事件隐式因果关系抽取方法
US20230092274A1 (en) Training example generation to create new intents for chatbots
US11797769B1 (en) Artificial intelligence system using hybrid technique for task-oriented dialog management
CN117575008A (zh) 训练样本生成方法、模型训练方法、知识问答方法及设备
CN108304513B (zh) 增加生成式对话模型结果多样性的方法及设备
CN111401069A (zh) 会话文本的意图识别方法、意图识别装置及终端
CN113704452B (zh) 基于Bert模型的数据推荐方法、装置、设备及介质
CN113535925B (zh) 语音播报方法、装置、设备及存储介质
CN112989044B (zh) 文本分类方法、装置、设备及存储介质

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: 21912790

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21912790

Country of ref document: EP

Kind code of ref document: A1