CN111091011B - Domain prediction method, domain prediction device and electronic equipment - Google Patents
Domain prediction method, domain prediction device and electronic equipment Download PDFInfo
- Publication number
- CN111091011B CN111091011B CN201911327989.6A CN201911327989A CN111091011B CN 111091011 B CN111091011 B CN 111091011B CN 201911327989 A CN201911327989 A CN 201911327989A CN 111091011 B CN111091011 B CN 111091011B
- Authority
- CN
- China
- Prior art keywords
- domain
- text
- round
- information
- field
- Prior art date
- Legal status (The legal status 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 status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 81
- 230000002452 interceptive effect Effects 0.000 claims abstract description 144
- 230000003993 interaction Effects 0.000 claims abstract description 114
- 238000009826 distribution Methods 0.000 claims abstract description 92
- 230000008569 process Effects 0.000 claims abstract description 27
- 230000014509 gene expression Effects 0.000 claims description 73
- 238000012549 training Methods 0.000 claims description 24
- 238000004590 computer program Methods 0.000 claims description 8
- 238000007781 pre-processing Methods 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 description 59
- 238000004891 communication Methods 0.000 description 12
- 238000013528 artificial neural network Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 5
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 5
- 230000006870 function Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000004091 panning Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/063—Training
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/08—Speech classification or search
- G10L15/16—Speech classification or search using artificial neural networks
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/08—Speech classification or search
- G10L15/18—Speech classification or search using natural language modelling
- G10L15/1822—Parsing for meaning understanding
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/22—Procedures used during a speech recognition process, e.g. man-machine dialogue
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Acoustics & Sound (AREA)
- Human Computer Interaction (AREA)
- General Engineering & Computer Science (AREA)
- Audiology, Speech & Language Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Machine Translation (AREA)
Abstract
The invention provides a field prediction method, a field prediction device and electronic equipment, wherein the field prediction method comprises the following steps: determining the interactive text of the round; the method comprises the steps of inputting a local interaction text and supervision information into a field prediction model to obtain field probability distribution which is output by the field prediction model and corresponds to the local interaction text, wherein the supervision information is field information which is determined after semantic understanding based on a previous interaction text, and correcting the field probability distribution which is output by the field prediction model and corresponds to the previous interaction text; and determining a domain prediction result based on the domain probability distribution corresponding to the interactive text of the round. The field prediction method provided by the embodiment of the invention can greatly improve the accuracy of model prediction in the multi-round interaction process, and particularly can obtain accurate field prediction results in the face of simplified interaction in the multi-round interaction process.
Description
Technical Field
The present invention relates to the field of voice interaction technologies, and in particular, to a field prediction method, a field prediction apparatus, and an electronic device.
Background
In the voice interaction process, in order to better understand semantics, it is generally necessary to predict which domain the user's expression content belongs to.
In the prior art, one method is to match whether the user expression content belongs to a certain field or not based on a grammar rule network or a state machine rule, and the generalization capability of the method is poor, and the sentence pattern which is not recorded can not be understood. The other mode is to use a deep neural network to learn sentence information of the expressed content of a user, so that the purpose of the model prediction field is achieved, but the accuracy of the deep neural network in the prediction of the expressed content of single-round interaction is still feasible, and once multiple rounds of dialogue are involved, the accuracy of the model prediction is greatly reduced.
Disclosure of Invention
Embodiments of the present invention provide a domain prediction method, a domain prediction apparatus, an electronic device, and a readable storage medium that overcome or at least partially solve the above-described problems.
In a first aspect, an embodiment of the present invention provides a domain prediction method, including: determining the interactive text of the round; inputting the local interactive text and the supervision information into a domain prediction model to obtain domain probability distribution corresponding to the local interactive text output by the domain prediction model, wherein the supervision information is determined based on the previous interactive text after semantic understanding, and the domain probability distribution corresponding to the previous interactive text output by the domain prediction model is corrected; determining a domain prediction result based on the domain probability distribution corresponding to the current interactive text; the field prediction model is obtained by training a plurality of rounds of interaction text data serving as samples in advance and the predetermined field probability distribution data corresponding to the rounds of interaction text data respectively serving as sample tags.
According to the field prediction method of the embodiment of the invention, the supervision information is obtained according to the following steps: acquiring the determined domain information of the previous interactive text after semantic understanding; inputting the previous round of interaction text into the domain prediction model to obtain domain probability distribution corresponding to the previous round of interaction text output by the domain prediction model; and determining the supervision information based on the domain information and the domain probability distribution corresponding to the previous round of interaction text.
According to the domain prediction method of the embodiment of the present invention, the step of inputting the current round of interaction text and the supervision information into a domain prediction model to obtain a domain probability distribution corresponding to the current round of interaction text output by the domain prediction model includes: inputting the local interactive text to a preprocessing layer of the domain prediction model to obtain local content characteristics and local domain word duty ratio characteristics, wherein the local content characteristics are used for representing the expression content of the local interactive text, and the local domain word duty ratio characteristics are used for representing the duty ratio of the length of each domain entity of the local interactive text to the local interactive text; and inputting the content characteristics of the current round, the duty ratio characteristics of the field words of the current round and the supervision information into an inference layer of the field prediction model to obtain the field probability distribution corresponding to the interactive text of the current round.
According to the domain prediction method of the embodiment of the present invention, the inputting the content feature of the present round, the duty ratio feature of the present round domain word and the supervision information to the inference layer of the domain prediction model, to obtain the domain probability distribution corresponding to the present round interaction text includes: inputting the content characteristics of the current round and the occupation bits of the field words of the current round into a first layer structure of the reasoning layer to obtain text expressions with field information occupation ratios and text expressions with field classification information, wherein the text expressions with field information occupation ratios are used for representing field information of the interactive text of the current round, and the text expressions with field classification information are used for representing predicted field occupation ratio weights of the interactive text of the current round; and inputting the text expression with the domain information ratio, the text expression with the domain classification information and the supervision information into a second layer structure of the reasoning layer to obtain the domain probability distribution corresponding to the interactive text of the round.
According to the domain prediction method provided by the embodiment of the invention, the text expression with the domain classification information is determined based on the content characteristics of the current round, the duty ratio characteristics of the current round of domain words and the weights of the domain classification information of the domain prediction model, and the weights of the domain classification information are determined according to each domain information learned by the domain prediction model in the training process.
According to the domain prediction method of the embodiment of the present invention, the inputting the text of the present round of interaction and the supervision information into the domain prediction model further includes: and inputting the text of the round of interaction, the supervision information and the personalized features into the field prediction model, wherein the personalized features are used for representing the associated information in the round of interaction.
According to the field prediction method provided by the embodiment of the invention, the personalized features comprise the foreground field features of the round of interaction and the background field features of the round of interaction.
In a second aspect, an embodiment of the present invention provides a domain prediction apparatus, including: the text determining unit is used for determining the interactive text of the round; the probability distribution determining unit is used for inputting the local interactive text and the supervision information into a field prediction model to obtain the field probability distribution which is output by the field prediction model and corresponds to the local interactive text, wherein the supervision information is the field information which is determined based on the previous interactive text after semantic understanding, and the field probability distribution which is output by the field prediction model and corresponds to the previous interactive text is corrected; the domain determining unit is used for determining a domain prediction result based on the domain probability distribution corresponding to the current interactive text; the field prediction model is obtained by training a plurality of rounds of interaction text data serving as samples in advance and a plurality of predetermined field probability distributions corresponding to the rounds of interaction text data respectively serving as sample tags.
In a third aspect, an embodiment of the present invention provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the domain prediction method as provided in the first aspect when the program is executed.
In a fourth aspect, embodiments of the present invention provide a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the domain prediction method as provided by the first aspect.
According to the domain prediction method, the domain prediction device, the electronic equipment and the readable storage medium, the domain prediction result of the round is predicted by using the domain probability distribution corresponding to the previous round of interactive text, the domain information determined after semantic understanding of the previous round of interactive text and the current round of interactive text, so that the current round of domain prediction considers the predicted domain of the previous round and the actual context of the previous round, the accuracy of model prediction in the multi-round interactive process is greatly improved, and particularly the accurate domain prediction result can be obtained in the face of simplified interaction in the multi-round interactive process.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a domain prediction method according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a topology structure of a domain prediction model in an application stage according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a topology structure of a domain prediction model in a training phase according to an embodiment of the present invention;
FIG. 4 is a schematic structural diagram of a domain prediction apparatus according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
In the prior art, the deep neural network does not completely solve the prediction of short spoken language in multiple rounds of conversations in the process of interaction. For example, when the user says "navigation goes to the science big news fly" for the first time, the currently used deep neural network can predict that the current request belongs to the map u navigation field; however, when the user then speaks "wanda", the "wanda" is input into the deep neural network, and since the current round of expression content has no strong sentence information, the analysis from the entity information alone cannot predict which field the current round of expression content belongs to, so the model has poor understanding capability on such problems.
The following describes a domain prediction method according to an embodiment of the present invention with reference to fig. 1 to 3, which can be used for a scenario of multi-round interaction.
The field prediction method of the embodiment of the invention comprises the following steps:
and step S100, determining the interactive text of the round.
In an actual implementation, determining the current round of interaction text may include: acquiring interactive voice information of the round; and transferring the round of interactive voice information into a round of interactive text.
It should be noted that, the multiple rounds of interactions represent multiple related communications between the user and the artificial intelligence, where the round of interactions includes the current user speaking or meaning expression in other forms, and corresponding feedback performed by the artificial intelligence after receiving the round of meaning expression, and correspondingly, the previous round of interactions is the round of interactions before the current round of interactions.
And step 200, inputting the interactive text of the current round and the supervision information into the field prediction model to obtain the field probability distribution corresponding to the interactive text of the current round, which is output by the field prediction model. The supervision information is domain information determined after semantic understanding based on the previous round of interaction text, and domain probability distribution corresponding to the previous round of interaction text output by the domain prediction model is corrected.
The domain information determined after semantic understanding of the previous interactive text refers to the domain information obtained after the final semantic understanding of the last expression content of the user. For example: the user speaks ' i want to listen to green fairy tale ', the model predicts that the domain of the current round of dialogue is the music domain, but after semantic understanding, the user's expression is understood as the story domain. Wherein the story domain is domain information after semantic understanding.
Taking the embodiment shown in fig. 2 as an example, the text input in the previous round of interaction is named as his_send, and the text input in the present round of interaction is named as send.
The above supervision information is denoted as his_attribute_vec, and the domain information determined by the previous round of interactive text after semantic understanding may be represented by a matrix his_label_vec, which may be determined by the domain matrix his_label_matrix of the previous round of interactive text after a series of semantic understanding.
For example, his_labelvec=his_labelmatrix T * The his_label_matrix, in other words, the transposed matrix of his_label_matrix is multiplied by his_label_matrix to obtain his_label_vec, so that the matrix dimension of his_label_vec becomes larger, and the information amount is rich.
The domain information determined after semantic understanding is used for correcting the domain probability distribution of the previous round, and the domain information (matrix his_label_vec) related to the semantic understanding result is fully connected with the domain probability distribution of the previous round to obtain a supervision matrix (supervision information), and the supervision matrix is used for supervising the domain prediction of the previous round.
When the domain prediction model is used to determine the domain probability distribution corresponding to the interactive text of the current round, the domain information (his_label_vec) determined after semantic understanding of the previous round of interactive text is used for supervising the domain probability distribution of the current round of interaction. For example, in the previous interactive text prediction process, when the predicted domain of the previous domain prediction model is consistent with the domain after semantic understanding of the previous round (assumed to be the music domain), the weight of the his_send_vec belonging to the music domain will be enhanced, and the weights of other domains will be weakened; when the domain predicted by the previous-round domain prediction model is inconsistent with the domain predicted by the previous-round semantic understanding (assuming that the domain prediction model predicts as the music domain and the domain actually understood as the story domain), the weight of the his_send_vec belonging to the story domain is enhanced, the weight of other domains is weakened, and when the prediction of the interactive text of the present round belongs to a certain domain, the music domain is not believed more, and the interactive text of the present round is believed to be the story domain more.
The field prediction model is obtained by training a plurality of rounds of interaction text data serving as samples in advance and the predetermined field probability distribution data corresponding to the rounds of interaction text data respectively serving as sample tags.
Taking fig. 3 as an example, during training, the previous round of interactive text data, the domain probability distribution data corresponding to the previous round of interactive text data, the present round of interactive text data and the domain probability distribution data corresponding to the present round of interactive text data are input, wherein the previous round of interactive text data and the present round of interactive text data are used as training samples, and the domain probability distribution data corresponding to the previous round of interactive text data and the domain probability distribution data corresponding to the present round of interactive text data are used as training sample labels.
Through training of a certain amount of sample data and sample labels, a trained field prediction model can be obtained. In the training stage, the domain prediction model does not need to input domain information (his_label_vec) or domain matrix (his_label_matrix) determined by the interactive text after semantic understanding.
And step S300, determining a domain prediction result based on the domain probability distribution corresponding to the interactive text of the round.
In a practical implementation, the domain probability distribution may be in the form of a probability distribution matrix in which the domain corresponding to the maximum value may be determined as the domain prediction result.
According to the field prediction method provided by the embodiment of the invention, the field prediction result of the current round is predicted by using the field probability distribution corresponding to the previous round of interactive text, the field information determined after semantic understanding of the previous round of interactive text and the current round of interactive text, so that the current round of field prediction considers the predicted field of the previous round and the actual context of the previous round, the accuracy of model prediction in the multi-round interactive process is greatly improved, and the accurate field prediction result can be obtained especially for the simplified interaction in the multi-round interactive process.
In some embodiments, as shown in fig. 2, the domain prediction method in the embodiment of the present invention further includes:
and S010, acquiring the field information determined after semantic understanding of the previous interactive text.
For example, a domain matrix his_label_matrix determined after semantic understanding of the previous round of interactive text is obtained, and domain information his_label_vec determined after semantic understanding of the previous round of interactive text is determined based on the domain matrix his_label_matrix.
The method comprises the steps that the dimension of a matrix of service coding output of the previous round of semantic understanding is 1 x label (short for his label matrix), and if the service of the previous round of semantic understanding is the navigation field, the parameter of the navigation field is set to be 1, and the other parameters are all 0; or may be set in reverse, that is, the navigation field is set to have a bit parameter of 0 and the other bits are all 1.
Wherein his_label_vec=his_label_matrix T * The his_label_matrix, in other words, the transposed matrix of his_label_matrix is multiplied by his_label_matrix to obtain his_label_vec, so that the matrix dimension of his_label_vec becomes larger, and the information amount is rich.
And S020, inputting the previous round of interactive text into the field prediction model to obtain the field probability distribution which is output by the field prediction model and corresponds to the previous round of interactive text.
In actual execution, the domain probability distribution corresponding to the previous round of interactive text can be determined by a domain prediction model in the current round of prediction, or the obtained domain probability distribution is stored in the previous round of prediction and is called in the current round of prediction.
Step S030, determining supervision information based on the domain information and the domain probability distribution corresponding to the previous interaction text.
In actual execution, the domain information is fully connected with the domain probability distribution corresponding to the previous interactive text, so that the supervision information can be obtained.
In some embodiments, step S200, inputting the current interaction text and the supervision information into the domain prediction model to obtain a domain probability distribution corresponding to the current interaction text output by the domain prediction model, includes:
And S210, inputting the interactive text of the current round into a preprocessing layer of the domain prediction model to obtain the content characteristics of the current round and the duty ratio characteristics of the words of the current round.
Taking the domain prediction model shown in fig. 2 as an example, the content of the present round is characterized by "forgetting water", and the domain word duty ratio of the present round is characterized by "domain word duty ratio".
The content features of the present round are used for representing the expression content of the interactive text of the present round, and the content features of the present round may be in the form of a word-embedding matrix, for example, the word-embedding matrix may be generated by a neural network or other network models.
The local domain word occupation ratio is used for representing the occupation ratio of the length of each domain entity of the local interactive text. Specifically, by sorting the domain vocabulary of each domain, the user's expressed content is unmatched to the domain vocabulary (there are many methods currently available for the algorithm for matching domain vocabulary, such as AC automaton (Aho-Corasick automaton)), and how many words in the currently expressed content belong to the domain vocabulary, so as to obtain the percentage of the number of the domain vocabulary in the whole expressed content.
The domain word duty ratio feature of the present round can be determined in a domain dictionary arc code form, wherein the matrix dimension output by the domain dictionary arc code is 1×label, the domain dictionary arc code refers to collecting and sorting all entities in a certain domain together, and the method of entity arc pasting (the implementation manner of entity arc pasting can be a regular matching method, an AC automaton method, etc., which are not described in detail herein) is used for carrying out arc pasting on the expression content of the user, so as to obtain the length ratio of the number of the entities belonging to the field in the expression content of the user to the text of the whole expression content.
And S220, inputting the content characteristics of the current round, the duty ratio characteristics of the words in the current round and the supervision information into an inference layer of the field prediction model to obtain the field probability distribution corresponding to the interactive text of the current round.
Taking fig. 2 as an example, the content feature (forgetting water), the domain word duty feature (domain word duty) and the supervision information (his_intent_vec) of the present round are input to the inference layer of the domain prediction model, so that domain probability distribution (domain) corresponding to the interactive text of the present round can be obtained.
When the inference layer predicts, the inference layer simultaneously considers the expression content of the interactive text of the round, the proportion of the length of each field entity of the interactive text of the round to the interactive text of the round and the supervision information of the interactive text of the previous round to predict the field probability distribution corresponding to the interactive text of the round, and when the simplified interaction in the multi-round interaction process is faced, the accurate field prediction result can be obtained.
In some embodiments, step S220, inputting the content feature of the present round, the duty ratio feature of the present round domain word and the supervision information to an inference layer of the domain prediction model, to obtain a domain probability distribution corresponding to the text of the present round of interaction, includes:
step S221, inputting the content characteristics of the current round and the occupation ratio of the words in the current round into a first layer structure of an inference layer to obtain text expressions with domain information occupation ratios and text expressions with domain classification information, wherein the text expressions with the domain information occupation ratios are used for representing the domain information of the current round of interactive texts, and the text expressions with the domain classification information are used for representing the prediction domain occupation ratio weights of the current round of interactive texts.
Taking the domain prediction model shown in fig. 2 as an example, the content feature (forgetting water) of the present round and the domain word duty feature (domain word duty) of the present round are input into the first layer structure of the inference layer, so as to obtain a text expression (send_vec) with the domain information duty and a text expression (text_label_vec) with the domain classification information.
Further, the reasoning layer includes: the pretreatment layer is used for determining the original text expression according to the content characteristics of the round; the first layer structure is used for determining the text expression with domain information duty ratio according to the original text expression and the current domain word duty ratio characteristic, and the second layer structure is used for determining the text expression with domain classification information according to the original text expression, the current domain word duty ratio characteristic and the weight of the domain classification information. The pretreatment layer can be a coding layer, the first layer structure can be an attention layer, and the attention processing can be carried out; the second layer structure may be a fully connected layer, and full connection may be performed.
Step S221, inputting the content features of the present round and the occupation bits of the words in the present round to the first layer structure of the inference layer, and obtaining the text expression with the occupation ratio of the domain information and the text expression with the classification information of the domain may include:
Step S221a, inputting the content characteristics of the round into a preprocessing layer of an inference layer to obtain the original text expression. In actual implementation, as shown in fig. 2, the original text expression (send_veco) is obtained after Bi-LSTM encoding the content feature (send, such as forgetting water) of the present round.
Step S221b, determining the text expression with domain information duty ratio according to the original text expression and the duty ratio characteristics of the domain words. In actual implementation, as shown in fig. 2, the original text expression (send_vec) and the current domain word occupation ratio feature (arc_vec) are subjected to the occupation to obtain the text expression (send_vec) with the domain information occupation ratio.
Step S221c, determining the text expression with the domain classification information according to the original text expression, the duty ratio characteristic of the field words of the round and the weight of the domain classification information. In actual implementation, as shown in fig. 2, the text expression (text_label_vec) with the domain classification information is obtained by performing the text expression (text_vec) on the original text expression (send_vec), the local domain word duty feature (arc_vec), and the weight (label_unbedding) of the domain classification information. Or, the original text expression (send_veco) and the local domain word duty ratio feature (arc_vec) are subjected to the attribute first, and the attribute result and the weight (label_unbedding) of the domain classification information are subjected to the attribute again to obtain the text expression (attribute_label_vec) with the domain classification information.
And step S222, inputting the text expression with the domain information proportion, the text expression with the domain classification information and the supervision information into a second layer structure of the reasoning layer to obtain the domain probability distribution corresponding to the interactive text of the round.
Taking the field prediction model shown in fig. 2 as an example, a text expression (send_vec) with the field information ratio, a text expression (text_label_vec) with the field classification information, and supervision information (his_text_vec) are input into a second layer structure of the inference layer, so as to obtain a field probability distribution (domain) corresponding to the interactive text of the present round.
Specifically, the text expression (send_vec) with the domain information duty ratio, the text expression (attribute_label_vec) with the domain classification information, and the supervision information (his_attribute_vec) are fully connected through the second layer structure:
fc_vec＝[attention_label_vec,sent_vec,his_attention_vec]，
fc＝weight*(fc_vec*fc_vec T )+b。
the text expression with the domain classification information is determined based on the content characteristics of the current round, the duty ratio characteristics of the words of the current round and the weights of the domain classification information of the domain prediction model, and the weights of the domain classification information are determined according to the domain information learned by the domain prediction model in the training process. In other words, the weight of the domain classification information may be that the domain prediction model remains after training is completed. Of course, the weight of the domain classification information may be determined in another model, and then the weight of the domain classification information determined by another model may be input to the domain prediction model.
In some preferred embodiments, step S200, inputting the text of the present round of interaction and the supervision information into the domain prediction model, further includes: and inputting the text, the supervision information and the personalized features of the round of interaction into a field prediction model, wherein the personalized features are used for representing the associated information in the round of interaction.
Compared with the embodiment, the method has the advantages that the personalized features are added during input so as to cope with the situation that basically the same multiple prediction results appear during field prediction, and the personalized features are added to assist the field prediction model in further distinguishing the ambiguous results.
Personalized features refer to information which can be extracted from the current interaction of the user according to different model application scenes and has certain association with the current interaction of the user, including but not limited to the states of other application devices in the foreground and background states of the whole interaction device in the process of using the interaction device by the user. For example, when the user uses the interaction device to interact with voice related to navigation, but at the moment, the background of the interaction device of the user also opens other application apps such as music, panning and the like; then, the navigation field belongs to the foreground, the music field and the shopping field belong to the background.
Further, the personalized features include foreground domain features of the present round of interactions and background domain features of the present round of interactions.
The matrix dimension of the personalized data code output of the personalized features is 1 x 2label (abbreviated as personal_matrix), and the 1 x 2label matrix is formed by splicing 2 1 x label matrices. The first 1 x label represents the foreground field in the current interaction, if the navigation field is in the foreground in the current interaction process, the navigation field's bit parameter is set to 1, and the other bit parameters are all 0. The second 1 x table represents the background field in the current interaction, and the parameters of all the fields in the background in the current interaction process are set to be 1, and the other parameters are all 0.
In some embodiments, step S220 is to input the content feature of the current round, the word duty feature of the current round, the personalized feature and the supervision information to an inference layer of the domain prediction model, so as to obtain a domain probability distribution corresponding to the interactive text of the current round.
Taking fig. 2 as an example, the current round of content features (forgetting water), the current round of domain word duty features (domain word duty), the personalized features (personal_matrix), and the supervision information (his_attribute_vec) are input to the reasoning layer of the domain prediction model, so that a domain probability distribution (domain) corresponding to the current round of interactive text can be obtained.
In step S222, the text expression with the domain information ratio, the text expression with the domain classification information, the personalized feature (personalized_matrix) and the supervision information are input into the second layer structure of the reasoning layer, so as to obtain the domain probability distribution corresponding to the interactive text of the present round.
Taking the domain prediction model shown in fig. 2 as an example, the personalized features (personalized_matrix) are converted into matrices per_vec by, but not limited to, directly adding a first matrix used to represent the foreground domain in the present round of interaction with a second matrix used to represent the background domain in the present round of interaction; or a first matrix used for representing the foreground field in the round of interaction and a second matrix used for representing the background field in the round of interaction are weighted and added.
Fully concatenating the text representation with field information (send_vec), the text representation with field classification information (personalized_label_vec), the personalized features (pers_vec) and the supervision information (his_personalized_vec),
fc_vec＝[attention_label_vec,sent_vec,pers_vec,his_attention_vec]
fc＝weight*(fc_vec*fc_vec T )+b，
and obtaining domain probability distribution (domain) corresponding to the interactive text of the round.
In one embodiment of the present invention, a domain prediction method includes: the method comprises the steps of inputting a previous round of interactive text and domain information determined by the previous round of interactive text after semantic understanding into a first part of a domain prediction model to obtain supervision information, wherein the supervision information is used for representing corrected domain probability distribution corresponding to the previous round of interactive text; inputting the text of the current round of interaction, personalized features and supervision information into a second part of the domain prediction model to obtain domain probability distribution corresponding to the text of the current round of interaction, wherein the personalized features are used for representing associated information in the current round of interaction; determining a domain prediction result based on domain probability distribution corresponding to the interactive text of the round; the field prediction model is obtained by training a sample label by taking multi-round interaction text data as a sample and taking a predetermined multi-round interaction field probability distribution corresponding to the multi-round interaction text data.
The scheme predicts the field of the expression content of the round jointly by using the expression content of the previous round of the user, the final semantic understanding field information of the expression content of the previous round and the personalized features which can be extracted in the interaction process of the round on the basis of fully using the generalization capability of the deep neural network. Therefore, the problem that the model prediction accuracy is not high in the multi-round interaction process in the existing scheme is solved.
The training process of the domain prediction model according to the embodiment of the present invention is described below with reference to fig. 3.
As shown in fig. 3, the left part is the text of the previous interaction, and the right part is the text of the present interaction; the text input in the previous round of interaction is named as his_send, and the text input in the current round of interaction is named as send.
And firstly, performing Bi-LSTM coding on a text his_send which is interacted with a user in the left part to obtain his_send_veco, and performing the attitution on the his_send_vec and his_arc_vec directly to obtain a sentence expression his_send_vec of an original text, wherein the his_send_vec matrix reflects the duty ratio of the domain information in the sentence of the original text.
his_sent_veco＝Bi-LSM(his_sent)
his_sent_vec＝attention(his_sent_veco,his_arc_vec)
And then performing the attitudes on his_send_vec, his_arc_vec and his_label_unbedding (his_label_unbedding is a matrix generated in the training process of the model, the matrix represents the information of each learned field when the model predicts, and the dimension of the matrix is label_256), calculating the feature matrix of the previous interactive text in the field, and generating his_attitudes_label_vec.
Finally, the his_attribute_label_vec and his_send_vec are fully connected, i.e
In order for the model to learn better about the previous round of interactive text weight, a penalty function is used above,
wherein y is i Representing the score of the previous interactive text in each field, wherein y is a matrix of 1 x label, and the higher the score of which matrix is, the higher the probability that the model considers that the previous interactive text prediction belongs to the field corresponding to the matrix is; y is i -representing the model actual output.
By using cross entropy loss functions, the results of model predictions are made to approach more and more the correct results as the model is trained.
As shown in fig. 3, a domain probability distribution (his_attribute_vec) corresponding to the previous round of interactive text is determined based on a text expression (his_send_vec) with a domain information duty ratio and a text expression (his_attribute_label_vec) with domain classification information.
In the right part, the processing logic is identical to that of the left part in the processing process of the interactive text of the round. After the attribute_label_vec matrix of the interaction of the round is obtained, full connection operation is carried out on his_attribute_vec, pers_vec and send_ vec, attention _label_vec of the text of the previous round of interaction. Namely:
fc_vec＝[attention_label_vec,sent_vec,pers_vec,his_attention_veco]
fc＝weight*(fc_vec*fc_vec T )+b
Similarly, in order for the model to learn the weight of the interactive text of the current round better, the same penalty function as the previous round of interaction is used,
wherein J i Representing the scores of the predicted interactive text of the round in various fields, J i -representing the model actual output.
In the training process, the domain probability distribution label his_domain_label corresponding to the previous round of interactive text needs to be input, and the domain probability distribution label domain_label corresponding to the current round of interactive text needs to be input.
The application process of the domain prediction model and the input network topology result of the model training process are slightly different, and the difference is that in the network topology structure of the previous round of interaction, a determined domain matrix his_label_matrix of the previous round of interaction text after semantic understanding is required to be input.
In the application network of the domain prediction model, when the supervision information his_attribute_vec is calculated, one his_label_vec matrix is added compared with the fact that his_attribute_vec is determined in the training model. When the actual model is tested, the field matrix of the last round of interactive text after a series of semantic understanding is his_label_matrix, and the specific his_label_vec calculation formula is as follows:
his_label_vec＝his_label_matrix T *his_label_matrix
his_attention_vec＝[his_sent_vec,his_attention_label_vec,his_label_vec]
fc_vec＝[attention_label_vec,sent_vec,pers_vec,his_attention_vec]
fc＝weight*(fc_vec*fc_vec T )+b
multiplying the transposed matrix of his_label_matrix by his_label_matrix to obtain the matrix dimension of his_label_vec, which is larger and the information quantity is rich. When the field prediction model predicts the field of the interactive text of the current round, the his_label_vec is used for supervising the model prediction result of the interactive text of the current round. For example, in the previous interactive text prediction process, when the predicted domain of the previous domain prediction model is consistent with the domain after semantic understanding of the previous round (assumed to be the music domain), the weight of the his_send_vec belonging to the music domain will be enhanced, and the weights of other domains will be weakened; when the domain predicted by the previous-round domain prediction model is inconsistent with the domain predicted by the previous-round semantic understanding (assuming that the model prediction is a music domain and the domain of actual semantic understanding is a story domain), the weight of the his_send_vec belonging to the story domain is enhanced, the weight of other domains is weakened, and when the prediction of the interactive text of the present round belongs to a certain domain, the music domain is not believed more, and instead the interactive text of the present round is believed to be the story domain more.
According to the field prediction method provided by the embodiment of the invention, on the basis of fully using the generalization capability of the deep neural network, the field of the expression content of the current round can be predicted jointly by using the expression content of the previous round of the user, the final semantic understanding field information of the expression content of the previous round and the personalized features which can be extracted in the interaction process of the current round, so that the problem that the model prediction accuracy of the existing scheme in the multi-round interaction process is not high can be solved.
The following describes a domain prediction apparatus provided by an embodiment of the present invention, and the domain prediction apparatus described below and the domain prediction method described above may be referred to correspondingly to each other.
As shown in fig. 4, the domain prediction apparatus according to the embodiment of the present invention includes: a text determining unit 710, configured to determine a text of the current round of interaction; the probability distribution determining unit 720 is configured to input the current round of interaction text and the supervision information into the domain prediction model to obtain a domain probability distribution corresponding to the current round of interaction text output by the domain prediction model, where the supervision information is determined based on the domain information determined after semantic understanding of the previous round of interaction text, and the domain probability distribution corresponding to the previous round of interaction text output by the domain prediction model is corrected to obtain the domain probability distribution; a domain determining unit 730, configured to determine a domain prediction result based on a domain probability distribution corresponding to the current interactive text; the field prediction model is obtained by training a plurality of rounds of interaction text data serving as samples in advance and a plurality of predetermined field probability distributions corresponding to the rounds of interaction text data respectively serving as sample tags.
Fig. 5 illustrates a physical schematic diagram of an electronic device, as shown in fig. 5, which may include: processor 810, communication interface (Communications Interface) 820, memory 830, and communication bus 840, wherein processor 810, communication interface 820, memory 830 accomplish communication with each other through communication bus 840. The processor 810 may invoke logic instructions in the memory 830 to perform a domain prediction method comprising: determining the interactive text of the round; inputting the local interactive text and the supervision information into a domain prediction model to obtain domain probability distribution corresponding to the local interactive text output by the domain prediction model, wherein the supervision information is determined based on the previous interactive text after semantic understanding, and the domain probability distribution corresponding to the previous interactive text output by the domain prediction model is corrected; determining a domain prediction result based on the domain probability distribution corresponding to the current interactive text; the field prediction model is obtained by training a plurality of rounds of interaction text data serving as samples in advance and the predetermined field probability distribution data corresponding to the rounds of interaction text data respectively serving as sample tags.
It should be noted that, in this embodiment, the electronic device may be a server, a PC, or other devices in the specific implementation, so long as the structure of the electronic device includes a processor 810, a communication interface 820, a memory 830, and a communication bus 840 as shown in fig. 5, where the processor 810, the communication interface 820, and the memory 830 complete communication with each other through the communication bus 840, and the processor 810 may call logic instructions in the memory 830 to execute the above method. The embodiment does not limit a specific implementation form of the electronic device.
Further, the logic instructions in the memory 830 described above may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
Further, embodiments of the present invention disclose a computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions which, when executed by a computer, enable the computer to perform the execution domain prediction method provided by the above method embodiments, the method comprising: determining the interactive text of the round; inputting the local interactive text and the supervision information into a domain prediction model to obtain domain probability distribution corresponding to the local interactive text output by the domain prediction model, wherein the supervision information is determined based on the previous interactive text after semantic understanding, and the domain probability distribution corresponding to the previous interactive text output by the domain prediction model is corrected; determining a domain prediction result based on the domain probability distribution corresponding to the current interactive text; the field prediction model is obtained by training a plurality of rounds of interaction text data serving as samples in advance and the predetermined field probability distribution data corresponding to the rounds of interaction text data respectively serving as sample tags.
In another aspect, embodiments of the present invention further provide a non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, is implemented to perform the execution domain prediction method provided in the above embodiments, the method including: determining the interactive text of the round; inputting the local interactive text and the supervision information into a domain prediction model to obtain domain probability distribution corresponding to the local interactive text output by the domain prediction model, wherein the supervision information is determined based on the previous interactive text after semantic understanding, and the domain probability distribution corresponding to the previous interactive text output by the domain prediction model is corrected; determining a domain prediction result based on the domain probability distribution corresponding to the current interactive text; the field prediction model is obtained by training a plurality of rounds of interaction text data serving as samples in advance and the predetermined field probability distribution data corresponding to the rounds of interaction text data respectively serving as sample tags.
The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (10)
1. A domain prediction method, comprising:
determining the interactive text of the round;
inputting the local interactive text and the supervision information into a domain prediction model to obtain domain probability distribution corresponding to the local interactive text output by the domain prediction model, wherein the supervision information is determined based on the previous interactive text after semantic understanding, and the domain probability distribution corresponding to the previous interactive text output by the domain prediction model is corrected; the supervision information is specifically obtained by fully connecting the domain information with domain probability distribution corresponding to the previous round of interaction text;
determining a domain prediction result based on the domain probability distribution corresponding to the current interactive text;
the field prediction model is obtained by taking multi-round interactive text data as a sample and taking pre-determined field probability distribution data corresponding to the multi-round interactive text data as a sample label.
2. The domain prediction method according to claim 1, wherein the supervision information is obtained by:
acquiring the determined domain information of the previous interactive text after semantic understanding;
Inputting the previous round of interaction text into the domain prediction model to obtain domain probability distribution corresponding to the previous round of interaction text output by the domain prediction model;
and determining the supervision information based on the domain information and the domain probability distribution corresponding to the previous round of interaction text.
3. The domain prediction method according to claim 1, wherein the step of inputting the native interaction text and the supervision information into a domain prediction model to obtain a domain probability distribution corresponding to the native interaction text output by the domain prediction model includes:
inputting the local interactive text to a preprocessing layer of the domain prediction model to obtain local content characteristics and local domain word duty ratio characteristics, wherein the local content characteristics are used for representing the expression content of the local interactive text, and the local domain word duty ratio characteristics are used for representing the duty ratio of the length of each domain entity of the local interactive text to the local interactive text;
and inputting the content characteristics of the current round, the duty ratio characteristics of the field words of the current round and the supervision information into an inference layer of the field prediction model to obtain the field probability distribution corresponding to the interactive text of the current round.
4. The domain prediction method according to claim 3, wherein the inputting the content feature of the present round, the duty ratio feature of the present round domain word, and the supervision information into the inference layer of the domain prediction model, to obtain the domain probability distribution corresponding to the present round interaction text, includes:
inputting the content characteristics of the current round and the occupation bits of the field words of the current round into a first layer structure of the reasoning layer to obtain text expressions with field information occupation ratios and text expressions with field classification information, wherein the text expressions with field information occupation ratios are used for representing field information of the interactive text of the current round, and the text expressions with field classification information are used for representing predicted field occupation ratio weights of the interactive text of the current round;
and inputting the text expression with the domain information ratio, the text expression with the domain classification information and the supervision information into a second layer structure of the reasoning layer to obtain the domain probability distribution corresponding to the interactive text of the round.
5. The domain prediction method according to claim 4, wherein the text expression of the domain-classified information is determined based on the current round content feature, the current round domain word duty feature, and the weight of the domain-classified information of the domain prediction model, and the weight of the domain-classified information is determined according to each domain information learned by the domain prediction model in a training process.
6. The domain prediction method according to any one of claims 1 to 5, wherein the inputting the present-round interaction text and supervision information into a domain prediction model includes:
and inputting the text of the round of interaction, the supervision information and the personalized features into the field prediction model, wherein the personalized features are used for representing the associated information in the round of interaction.
7. The domain prediction method according to claim 6, wherein the personalized features include foreground domain features of the present round of interaction and background domain features of the present round of interaction.
8. A domain prediction device, comprising:
the text determining unit is used for determining the interactive text of the round;
the probability distribution determining unit is used for inputting the local interactive text and the supervision information into a field prediction model to obtain the field probability distribution which is output by the field prediction model and corresponds to the local interactive text, wherein the supervision information is the field information which is determined based on the previous interactive text after semantic understanding, and the field probability distribution which is output by the field prediction model and corresponds to the previous interactive text is corrected; the supervision information is specifically obtained by fully connecting the domain information with domain probability distribution corresponding to the previous round of interaction text;
The domain determining unit is used for determining a domain prediction result based on the domain probability distribution corresponding to the current interactive text; wherein the method comprises the steps of
The field prediction model is obtained by training a plurality of rounds of interaction text data serving as samples in advance and a plurality of predetermined field probability distributions corresponding to the rounds of interaction text data respectively serving as sample tags.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the domain prediction method of any one of claims 1 to 8 when the program is executed by the processor.
10. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the domain prediction method of any of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911327989.6A CN111091011B (en) | 2019-12-20 | 2019-12-20 | Domain prediction method, domain prediction device and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911327989.6A CN111091011B (en) | 2019-12-20 | 2019-12-20 | Domain prediction method, domain prediction device and electronic equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111091011A CN111091011A (en) | 2020-05-01 |
| CN111091011B true CN111091011B (en) | 2023-07-28 |
Family
ID=70395878
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911327989.6A Active CN111091011B (en) | 2019-12-20 | 2019-12-20 | Domain prediction method, domain prediction device and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111091011B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111639160A (en) * | 2020-05-29 | 2020-09-08 | 达闼机器人有限公司 | Domain identification method, interaction method, electronic device and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106530305A (en) * | 2016-09-23 | 2017-03-22 | 北京市商汤科技开发有限公司 | Semantic segmentation model training and image segmentation method and device, and calculating equipment |
| CN109960749A (en) * | 2019-02-22 | 2019-07-02 | 清华大学 | Model acquisition methods, keyword generation method, device, medium and calculating equipment |
| WO2019136946A1 (en) * | 2018-01-15 | 2019-07-18 | 中山大学 | Deep learning-based weakly supervised salient object detection method and system |
Family Cites Families (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7519566B2 (en) * | 2004-02-11 | 2009-04-14 | Oracle International Corporation | Method and apparatus for automatically and continuously updating prediction models in real time based on data mining |
| US9792281B2 (en) * | 2015-06-15 | 2017-10-17 | Microsoft Technology Licensing, Llc | Contextual language generation by leveraging language understanding |
| CN107632987B (en) * | 2016-07-19 | 2018-12-07 | 腾讯科技（深圳）有限公司 | A kind of dialogue generation method and device |
| CN108153780B (en) * | 2016-12-05 | 2021-11-23 | 阿里巴巴集团控股有限公司 | Man-machine conversation device and method for realizing man-machine conversation |
| CN106649739B (en) * | 2016-12-23 | 2020-09-11 | 广东惠禾科技发展有限公司 | Multi-round interactive information inheritance identification method and device and interactive system |
| CN106776578B (en) * | 2017-01-03 | 2020-03-17 | 竹间智能科技（上海）有限公司 | Method and device for improving conversation performance of conversation system |
| US10747954B2 (en) * | 2017-10-31 | 2020-08-18 | Baidu Usa Llc | System and method for performing tasks based on user inputs using natural language processing |
| WO2019147804A1 (en) * | 2018-01-26 | 2019-08-01 | Ge Inspection Technologies, Lp | Generating natural language recommendations based on an industrial language model |
| US20190294973A1 (en) * | 2018-03-23 | 2019-09-26 | Google Llc | Conversational turn analysis neural networks |
| CN109062937B (en) * | 2018-06-15 | 2019-11-26 | 北京百度网讯科技有限公司 | The method of training description text generation model, the method and device for generating description text |
| CN109582767B (en) * | 2018-11-21 | 2024-05-17 | 北京京东尚科信息技术有限公司 | Dialogue system processing method, device, equipment and readable storage medium |
| CN109857843B (en) * | 2018-12-25 | 2023-01-17 | 科大讯飞股份有限公司 | Interaction method and system based on document |
| CN109977209A (en) * | 2019-03-22 | 2019-07-05 | 深圳狗尾草智能科技有限公司 | More wheel man-machine interaction methods, system, computer and medium |
| CN110188204B (en) * | 2019-06-11 | 2022-10-04 | 腾讯科技（深圳）有限公司 | Extended corpus mining method and device, server and storage medium |
| CN110209791B (en) * | 2019-06-12 | 2021-03-26 | 百融云创科技股份有限公司 | Multi-round dialogue intelligent voice interaction system and device |
-
2019
- 2019-12-20 CN CN201911327989.6A patent/CN111091011B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106530305A (en) * | 2016-09-23 | 2017-03-22 | 北京市商汤科技开发有限公司 | Semantic segmentation model training and image segmentation method and device, and calculating equipment |
| WO2019136946A1 (en) * | 2018-01-15 | 2019-07-18 | 中山大学 | Deep learning-based weakly supervised salient object detection method and system |
| CN109960749A (en) * | 2019-02-22 | 2019-07-02 | 清华大学 | Model acquisition methods, keyword generation method, device, medium and calculating equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111091011A (en) | 2020-05-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20230028944A1 (en) | Dialogue generation method and network training method and apparatus, storage medium, and device | |
| CN109582767A (en) | Conversational system processing method, device, equipment and readable storage medium storing program for executing | |
| CN111191450B (en) | Corpus cleaning method, corpus input device and computer readable storage medium | |
| US11636272B2 (en) | Hybrid natural language understanding | |
| CN110069612B (en) | Reply generation method and device | |
| CN111339781A (en) | Intention recognition method and device, electronic equipment and storage medium | |
| CN108710704A (en) | Determination method, apparatus, electronic equipment and the storage medium of dialogue state | |
| CN112084317A (en) | Method and apparatus for pre-training a language model | |
| CN115497465A (en) | Voice interaction method and device, electronic equipment and storage medium | |
| CN111402864A (en) | Voice processing method and electronic equipment | |
| CN111091011B (en) | Domain prediction method, domain prediction device and electronic equipment | |
| CN113569017A (en) | Model processing method and device, electronic equipment and storage medium | |
| CN116186259A (en) | Session cue scoring method, device, equipment and storage medium | |
| CN113535930B (en) | Model training method, device and storage medium | |
| CN113761874A (en) | Event reality prediction method and device, electronic equipment and storage medium | |
| CN116776870B (en) | Intention recognition method, device, computer equipment and medium | |
| CN117453899B (en) | Intelligent dialogue system and method based on large model and electronic equipment | |
| CN115934920B (en) | Model training method for man-machine conversation and related device | |
| CN117453895B (en) | Intelligent customer service response method, device, equipment and readable storage medium | |
| CN117436457B (en) | Irony identification method, irony identification device, computing equipment and storage medium | |
| CN112685558B (en) | Training method and device for emotion classification model | |
| CN109241539B (en) | Updating method of machine learning artificial intelligence translation database | |
| CN114218939A (en) | Text word segmentation method, device, equipment and storage medium | |
| CN115617997A (en) | Dialog state tracking method, device, equipment and medium | |
| CN116561248A (en) | Information processing method and device and electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |