CN111160512A - Method for constructing dual-discriminator dialog generation model based on generative confrontation network - Google Patents
Method for constructing dual-discriminator dialog generation model based on generative confrontation network Download PDFInfo
- Publication number
- CN111160512A CN111160512A CN201911224148.2A CN201911224148A CN111160512A CN 111160512 A CN111160512 A CN 111160512A CN 201911224148 A CN201911224148 A CN 201911224148A CN 111160512 A CN111160512 A CN 111160512A
- Authority
- CN
- China
- Prior art keywords
- model
- rewriting
- discriminator
- reply
- context
- 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.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention discloses a method for constructing a dual-discriminator dialogue generating model based on a generating countermeasure network, which comprises the following steps of firstly processing corpora to obtain a quaternary corpora with similar dialogue information; then, a rewriting model and a discrimination model are pre-trained, wherein the former can rewrite the matched similar reply to generate a reply more in line with the current context, and the latter can discriminate whether the sentence is from a corpus or the rewriting model; and finally, the rewriting model and the discrimination model are used for counterstudy, and the best rewriting effect is obtained in the game process of the rewriting model and the discrimination model. The invention improves the generation model by introducing two discriminators from multiple angles, and makes great progress in the aspects of grammar, context correlation and the like of the generated sentence.
Description
Technical Field
The invention relates to a natural language processing, deep learning and dialogue system, in particular to a method for constructing a dual-discriminator dialogue generating model based on a generative countermeasure network (GAN).
Background
With the development of smart phones and smart homes, interaction between people and machines becomes more and more frequent, the requirement of users for the quality of conversation with the machines becomes higher and higher, and smooth, fluent and various communication experiences are expected to be obtained, which means that the template-based conversation system commonly used in the industry at present is difficult to meet the requirements of users. At present, the method for constructing a dialog system commonly used in the industry is based on templates, namely, manually arranging and defining a large number of dialogs to form the templates, and inputting the dialogs spoken by a user into the pre-defined templates to obtain a fixed reply, and the method covers fewer topics and has higher labor cost. With the development of big data technology, deep learning technology and computer hardware, the academic world explores a lot in the aspect of automatically generating a dialogue system, the big data technology provides a lot of analyzable corpora, the deep learning technology provides a complex computing frame, the computer hardware technology provides high-speed operation, and the three act together to promote the development of the dialogue system.
The generation type dialogue system is characterized in that a user obtains a reply which is created by the system, no one tells the system a fixed answer in advance, and only the user needs to learn by himself when losing a large amount of data, and the process of learning by himself uses deep learning. Most of the current generation-type dialogue systems are self-learning based on a seq2seq model, so that the model learns how to encode input and then decode the input to obtain a reply, and the model is improved by continuously reducing the difference between a generated statement and a real statement, however, the model is lazy, which only learns simple generation, namely, the model is more prone to generating ubiquitous replies of 'I do not know', 'good' and 'kayi'. Meanwhile, because the corpora adopted by the traditional methods are binary corpora such as 'above + reply', there is no intervention from reply sentences in the generation process, so that the model does not know what sentences are correct, and some utterances with wrong grammar and blind semantics are generated sometimes. Therefore, a method for improving the diversity, context correlation and grammar accuracy of the automatic dialog generation system is needed.
Disclosure of Invention
The invention aims to provide a method for constructing a dual-discriminator dialogue generating model based on a generative confrontation network aiming at the defects of the existing model of a generative dialogue system.
The specific technical scheme for realizing the aim of the invention is as follows:
a method for constructing a dual-discriminator dialogue generating model based on a generative confrontation network comprises the following specific steps:
step 1: corpus processing
Matching similar contexts C 'in a corpus by using a text matching algorithm according to the current context C, thereby obtaining a reply R' under the similar contexts to form a quaternary corpus < C, R, C 'and R' >;
step 2: initializing rewrite models
Training the quaternary data obtained in the step 1 in a seq2seq frame to obtain a primary rewriting model, wherein the rewriting model can rewrite R 'by combining the contexts C and C' to generate a reply R, and the rewriting model does not achieve an ideal effect and has a large loss;
and step 3: initialization discriminant model
And (3) training two discriminators of the discrimination model by using the reply R generated in the step (2) and the real reply R, wherein the two discriminators respectively judge from two aspects of rewriting effect and context correlation, and the specific process is as follows:
the discriminator _1 aims at discriminating the quality of the rewriting effect, the input of the discriminator is sentences before and after rewriting, wherein the sentence type before rewriting is "True", the sentence type after rewriting is "False", and the discriminator _1 is used for distinguishing two types as much as possible;
the discriminator _2 is used for discriminating whether the context correlation is good or bad, so that the other inputs are 'current context + generated reply' and 'current context + real reply', obviously, the context correlation of the latter is strongest, the category is marked as 'True', the former is marked as 'False', and the discriminator _2 is used for distinguishing the two categories as far as possible;
and 4, step 4: adapting model and discriminating model countertraining
The rewriting model updates parameters according to feedback of the discrimination model, then, the new rewriting model transmits the generated sentences to the discrimination model, the discrimination model updates the model parameters by using the accuracy rate of distinguishing true and false sentences, in the process of countermeasure, loss of the two sentences is in a descending trend until the sentences tend to be stable, and the countermeasure training is terminated; and the obtained rewriting model is an optimal model, namely the conversation generation model.
In the process of the confrontation training, the lower the accuracy of the discriminator is, the better the effect of rewriting the model is, because the discriminator is 'confused'; however, the discriminators also increase their discrimination power with increasing overwrite model effect and decreasing their own accuracy, which is a so-called "countermeasure" process. When the loss of the generated model and the accuracy of the discriminant model are reduced to a certain degree and tend to be stable, the two models are considered to be equivalent in strength, and the confrontation can be stopped. The generated model obtained at this time is the optimal model, and the rewritten model is taken as the final dialogue generated model.
Compared with the prior art, the invention has the following advantages:
1) the grammar is correct: compared with the method of generating a sentence from zero, the method has the advantages that the method is rewritten based on the existing sentence, and a good grammar basis can be provided;
2) semantic smoothing: the method adapts according to replies in similar contexts, which can have a better context basis, so that the replies obtained are more context-compliant.
3) The model has strong automatic learning performance: and modifying the model parameters according to the effect of the opposite party, not only according to the loss of the model, by modifying the counterstudy of the model and the discriminator.
Drawings
Fig. 1 is an overall frame diagram of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the following detailed description and accompanying drawings. The procedures, conditions, experimental methods and the like for carrying out the present invention are general knowledge and common general knowledge in the art, and the present invention is not particularly limited, except for those specifically mentioned below.
Examples
Referring to fig. 1, the present invention provides a method for constructing a dual-discriminator dialog generation model based on a generative confrontation network, that is, a method for constructing a dialog generation model by replacing creation and rewriting models with rewriting and discriminator game learning, as shown in the figure, a current context C is known, and a similar context C 'and a reply R' thereof are obtained by matching using a text matching algorithm. The left dashed frame part in the figure is the rewrite model of this embodiment, which is based on the seq2seq frame, the encoder encodes R ', and the decoder decodes while introducing the difference diff (C, C ') between C and C ', and finally obtains the generated reply R. The right dashed frame part in the figure is the discrimination model of the embodiment, the discriminator _1 learns to distinguish R and true reply R obtained by the rewriting model, the discriminator _2 learns to distinguish false dialog C + R and true dialog C + R, and the feedback obtained by the two discriminators is returned to the rewriting model. The embodiment specifically comprises the following steps:
step 1: data pre-processing
Based on observation and experience, it is easier to modify on a template basis than to author from scratch, so this embodiment proposes to get a dialog reply by rewriting a sentence, and all that is needed is to process the corpus. The existing dialogue corpora are paired, that is, the current context (context) corresponds to a sentence reply (response), in order to have a better rewriting foundation, the two contexts are ensured to be under the condition of similar contexts, so that the similar contexts are obtained according to a matching algorithm, the corresponding replies are extracted to form a quaternary corpus < C, R, C ', R ' >, and the embodiment directly calls a text matching algorithm to obtain the first 10C's with the highest matching score with C to form the required corpora. After the quaternary linguistic data are obtained, the data are further cleaned, the maximum sentence length is set to be 50 words, and the words exceeding the threshold value are discarded. Further, the corpus is divided into a training set, a verification set and a test set, and the ratio is 7: 2: 1.
Step 2: pre-training rewrite model
The rewriting model is based on a common frame-seq 2seq of the generation model, the input of the frame is R ', the target output is R, the frame generates R word by word, and an Attention mechanism is introduced in the decoding process, wherein the Attention mechanism is from edit vectors diff (C, C'), R 'and the generated word sequence, and the edit vectors are the difference word vector sequences of C and C'. And (3) training the model by using the training set obtained in the step (1), obtaining a model after each training, verifying by using the verification set obtained in the step (1), and selecting the model with the best effect as an initialization model for rewriting the model in the countermeasure process.
And step 3: pre-training discrimination model
The arbiter is built based on a neural network and functions to determine whether a utterance/set of dialogs is genuine. Before a sentence is input into a discriminator, firstly, the sentence is labeled, the label from a corpus is 1, and the label generated by a rewriting model is 0; the discriminator is then trained to distinguish between the two classes of sentences as much as possible. To distinguish the sentence from the other sentence, the embodiment applies two classifiers, whose inputs are shown in fig. 1, and the discriminative ability of the classifier _1 for sentence grammatical is trained by using R and R', and the discriminative ability of the classifier _2 for context is trained by using C + R and C + R.
And 4, step 4: adapting model and discriminating model counterstudy
In the counterstudy process, the objective of the rewrite model is to generate a true answer to deceive the discriminant model, and the objective of the discriminant model is to distinguish the reply generated by the rewrite model from the true reply. Thus, the rewrite model and the discriminant model constitute a dynamic countermeasure. In the most ideal state, rewriting the model G may generate enough replies to be "spurious". For the discriminant model D, it is difficult to determine whether the recovery generated by the overwrite model is true or not, and the accuracy of the discriminant is D (g) = 0.5. In the process of confrontation, the parameters are continuously adjusted to make the accuracy of the discrimination model approach 0.5, and simultaneously, the loss of the rewriting model is reduced and finally tends to be stable. The generated reply of the finally obtained dialogue generation model is improved in grammar and context, and results of 0.629, 0.755 and 0.682 are respectively obtained on vector cosine similarity measurement indexes Greeny, Average and Extramem. This results in a generative model that can be used to generate the appropriate reply.
Claims (1)
1. A method for constructing a dual-discriminator dialogue generating model based on a generative confrontation network is characterized by comprising the following specific steps:
step 1: corpus processing
Matching similar contexts C 'in a corpus by using a text matching algorithm according to the current context C, thereby obtaining a reply R' under the similar contexts to form a quaternary corpus < C, R, C 'and R' >;
step 2: initializing rewrite models
Training the quaternary data obtained in the step 1 in a seq2seq frame to obtain a primary rewriting model, wherein the rewriting model can rewrite R 'by combining the contexts C and C' to generate a reply R;
and step 3: initialization discriminant model
And (3) training two discriminators of the discrimination model by using the reply R generated in the step (2) and the real reply R, wherein the two discriminators respectively judge from two aspects of rewriting effect and context correlation, and the specific process is as follows:
a discriminator _1 for discriminating the rewriting effect, wherein the input is the sentence before and after rewriting, the sentence before rewriting is "True", the sentence after rewriting is "False", and the discriminator _1 distinguishes two categories;
the discriminator _2, aiming at discriminating context correlation, inputs of which are "current context + generated reply" and "current context + real reply", the latter having the strongest context correlation, the category is marked as "True", the former is marked as "False", and the discriminator _2 distinguishes two categories;
and 4, step 4: adapting model and discriminating model countertraining
The rewriting model updates parameters according to feedback of the discrimination model, the updated rewriting model transmits the generated sentences to the discrimination model, the discrimination model updates the model parameters by using the accuracy rate of distinguishing true and false sentences, in the process of countermeasure, loss of the two sentences is in a descending trend until the sentences tend to be stable, and the countermeasure training is terminated; and the obtained rewriting model is an optimal model, namely the conversation generation model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911224148.2A CN111160512B (en) | 2019-12-04 | 2019-12-04 | Method for constructing double-discriminant dialogue generation model based on generation type countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911224148.2A CN111160512B (en) | 2019-12-04 | 2019-12-04 | Method for constructing double-discriminant dialogue generation model based on generation type countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111160512A true CN111160512A (en) | 2020-05-15 |
| CN111160512B CN111160512B (en) | 2023-06-13 |
Family
ID=70556431
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911224148.2A Active CN111160512B (en) | 2019-12-04 | 2019-12-04 | Method for constructing double-discriminant dialogue generation model based on generation type countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111160512B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111783955A (en) * | 2020-06-30 | 2020-10-16 | 北京市商汤科技开发有限公司 | Neural network training method, neural network training device, neural network dialogue generating method, neural network dialogue generating device, and storage medium |
| CN113076078A (en) * | 2021-03-11 | 2021-07-06 | 山东大学 | Hybrid drive-based conversational information acquisition method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004102596A (en) * | 2002-09-09 | 2004-04-02 | Mitsubishi Electric Corp | Language understanding model generating device |
| US20160350655A1 (en) * | 2015-05-26 | 2016-12-01 | Evature Technologies (2009) Ltd. | Systems Methods Circuits and Associated Computer Executable Code for Deep Learning Based Natural Language Understanding |
| CN108734276A (en) * | 2018-04-28 | 2018-11-02 | 同济大学 | A kind of learning by imitation dialogue generation method generating network based on confrontation |
| US10152970B1 (en) * | 2018-02-08 | 2018-12-11 | Capital One Services, Llc | Adversarial learning and generation of dialogue responses |
| CN110188182A (en) * | 2019-05-31 | 2019-08-30 | 中国科学院深圳先进技术研究院 | Model training method, dialogue generation method, device, equipment and medium |
| CN110347792A (en) * | 2019-06-25 | 2019-10-18 | 腾讯科技（深圳）有限公司 | Talk with generation method and device, storage medium, electronic equipment |
-
2019
- 2019-12-04 CN CN201911224148.2A patent/CN111160512B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004102596A (en) * | 2002-09-09 | 2004-04-02 | Mitsubishi Electric Corp | Language understanding model generating device |
| US20160350655A1 (en) * | 2015-05-26 | 2016-12-01 | Evature Technologies (2009) Ltd. | Systems Methods Circuits and Associated Computer Executable Code for Deep Learning Based Natural Language Understanding |
| US10152970B1 (en) * | 2018-02-08 | 2018-12-11 | Capital One Services, Llc | Adversarial learning and generation of dialogue responses |
| CN108734276A (en) * | 2018-04-28 | 2018-11-02 | 同济大学 | A kind of learning by imitation dialogue generation method generating network based on confrontation |
| CN110188182A (en) * | 2019-05-31 | 2019-08-30 | 中国科学院深圳先进技术研究院 | Model training method, dialogue generation method, device, equipment and medium |
| CN110347792A (en) * | 2019-06-25 | 2019-10-18 | 腾讯科技（深圳）有限公司 | Talk with generation method and device, storage medium, electronic equipment |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111783955A (en) * | 2020-06-30 | 2020-10-16 | 北京市商汤科技开发有限公司 | Neural network training method, neural network training device, neural network dialogue generating method, neural network dialogue generating device, and storage medium |
| CN113076078A (en) * | 2021-03-11 | 2021-07-06 | 山东大学 | Hybrid drive-based conversational information acquisition method |
| CN113076078B (en) * | 2021-03-11 | 2022-03-22 | 山东大学 | Hybrid drive-based conversational information acquisition method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111160512B (en) | 2023-06-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110781680B (en) | Semantic similarity matching method based on twin network and multi-head attention mechanism | |
| CN108710704B (en) | Method and device for determining conversation state, electronic equipment and storage medium | |
| CN109977207A (en) | Talk with generation method, dialogue generating means, electronic equipment and storage medium | |
| CN112669841A (en) | Training method and device for multilingual speech generation model and computer equipment | |
| CN110059169B (en) | Intelligent robot chat context implementation method and system based on corpus labeling | |
| Udagawa et al. | A natural language corpus of common grounding under continuous and partially-observable context | |
| CN111930918B (en) | Cross-modal bilateral personalized man-machine social interaction dialog generation method and system | |
| CN112037773B (en) | N-optimal spoken language semantic recognition method and device and electronic equipment | |
| CN115292463B (en) | Information extraction-based method for joint multi-intention detection and overlapping slot filling | |
| CN112487139A (en) | Text-based automatic question setting method and device and computer equipment | |
| CN112992125B (en) | Voice recognition method and device, electronic equipment and readable storage medium | |
| CN111754992A (en) | Noise robust audio/video bimodal speech recognition method and system | |
| CN110942774A (en) | Man-machine interaction system, and dialogue method, medium and equipment thereof | |
| CN113408535A (en) | OCR error correction method based on Chinese character level characteristics and language model | |
| CN111160512A (en) | Method for constructing dual-discriminator dialog generation model based on generative confrontation network | |
| CN114911932A (en) | Heterogeneous graph structure multi-conversation person emotion analysis method based on theme semantic enhancement | |
| CN113076749A (en) | Text recognition method and system | |
| CN115104151A (en) | Offline voice recognition method and device, electronic equipment and readable storage medium | |
| CN116978367A (en) | Speech recognition method, device, electronic equipment and storage medium | |
| CN114626529B (en) | Natural language reasoning fine tuning method, system, device and storage medium | |
| CN113486160B (en) | Dialogue method and system based on cross-language knowledge | |
| CN114463688A (en) | Cross-modal context coding dialogue emotion recognition method and system | |
| CN114358019A (en) | Method and system for training intention prediction model | |
| CN113362815A (en) | Voice interaction method, system, electronic equipment and storage medium | |
| CN115394288B (en) | Language identification method and system for civil aviation multi-language radio land-air conversation |
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 |