CN112131367A - Self-auditing man-machine conversation method, system and readable storage medium - Google Patents

Abstract

The present invention provides a self-audit man-machine dialogue method, system and readable storage medium. The self-audit man-machine dialogue system is composed of an input and output module, a dialogue reply generation module and a dialogue self-audit module. The user can enter any dialogue, generate multiple candidate responses through the generation model in the response generation module, and then calculate the loss of the candidate responses relative to the historical dialogue through the MMI scoring model, and select the one with the smallest loss as the final response. Input the generated reply into the dialogue self-audit model, first use the pre-trained Bert model to extract text features, and then use the bidirectional LSTM model for classification to determine whether it is a sensitive sentence, if not, output the reply to the user; If it is a sensitive sentence, the reply in the dialog template set in advance is output. The inventive system can be applied to any scene for chatting.

Description

Self-audit man-machine dialogue method, system and readable storage medium

【Technical field】

The present invention relates to the technical field of artificial intelligence processing, in particular to a self-audit man-machine dialogue method, system and readable storage medium.

【Background technique】

In recent years, with the continuous development of science and technology, dialogue robots have attracted the attention of scientific researchers, and have been widely used in various fields because of their huge commercial value and potential. A series of chat robot products have appeared, such as Microsoft Xiaoice, Ali Xiaomi, etc. The goal of conversational bots is to make any open-domain response to user-typed utterances. The rapid development of artificial intelligence and deep learning technology has greatly promoted the progress of dialogue systems. Generating human-like replies is no longer an illusion, and gradually allows the establishment of data-driven and open dialogue systems between humans and computers.

At present, the dialogue systems in the market can be divided into retrieval methods and generative methods. The retrieval method first expresses the entire context as a vector, and then performs semantic matching between the context vector and the vector of candidate responses in the corpus, and selects the most similar semantically to reply. However, this matching method is prone to lose contextual information, and the semantic gap between dialogue and reply still exists, and existing algorithms cannot overcome this problem well. The generative method mainly relies on a large amount of corpus, uses the neural network model to fully encode the dialogue context information, and trains a more "intelligent" dialogue robot. Pleasant conversation experience.

Therefore, it is necessary to study a self-auditing human-computer dialogue method, system and readable storage medium to deal with the deficiencies of the prior art, so as to solve or alleviate one or more of the above problems.

[Content of the invention]

In view of this, the present invention provides a self-audit man-machine dialogue method, system and readable storage medium, which can provide free and open dialogue in multiple fields, complete multiple rounds of dialogue with users on a topic, and can The content of the reply is reviewed to determine whether it is swearing or sensitive words, and simulate the human thinking mode, which is equivalent to restraining the system from a moral level, avoiding sending sensitive words to users and causing a bad experience, and always maintaining appropriate output in the dialogue scene. Comfortable words make customers have a pleasant chat experience.

On the one hand, the present invention provides a self-audit man-machine dialogue method, and the self-audit man-machine dialogue method specifically includes the following steps:

S1: Collect multiple different types of corpus to form a corpus;

S2: Use the GPT2 model to train the corpus, and splicing each round of dialogue contained in the corpus in the order in which multiple rounds of dialogue are generated;

S3: Encode the input language dialogue;

S4: Generate a candidate set consisting of multiple replies through the results of S2 and S3;

S5: Score the candidate set through the MMI model, and select the sentence with the smallest loss relative to the dialogue history in the candidate set as the final reply;

S6: Perform feature extraction and model classification on the final reply, and determine whether the final reply contains sensitive words. If it contains sensitive words, output a specific sentence. If it does not contain sensitive words, use the final reply as the output language dialogue.

The above aspect and any possible implementation manner further provide an implementation manner, the S3 is specifically: converting the input language dialogue into a vector using byte pair encoding, and expressing it as a long text x ₁ , x ₂ , ·, x _n , and a closing tag is set at the end.

The above-mentioned aspect and any possible implementation manner further provide an implementation manner, and the S4 specifically includes the following steps:

S41: Taking the input language dialogue as the source sentence, expressed as S=x ₁ , x ₂ , ·, x _m , where S is the source sentence;

S42: Use the replies in the corpus as the target sentence, expressed as T=x _m+1 , x ₂ , ·, x _n , where T is the target sentence, and the conditional probability of the target sentence is:

S43: Generate multiple candidate sentences through the autoregressive language model. The autoregressive language model uses the density estimation of the joint probability of the text sequence to model, and the formula is as follows:

S44: Select a plurality of generated candidate sentences through the Top-k Sampling method to form a candidate set composed of multiple replies.

The above-mentioned aspect and any possible implementation manner further provide an implementation manner, and the S5 specifically includes the following steps:

S51: splicing each candidate sentence in the candidate set with the dialogue history in reverse order;

S52: Taking each candidate sentence as the target, for the given reply hypothesis H of each candidate sentence, predict the source sentence S, and the calculation formula of the MMI model is as follows:

Among them, logp(H) represents the penalty item, and the penalty item is that the probability of the language model for the frequently occurring reply H in the training set will be higher than that of other replies, and subtracting the penalty item can reduce the score of the frequently occurring H;

S53: Select and penalize plain and general replies through the MMI model, calculate the loss of all candidate sentences relative to the dialogue history, and finally select the one with the smallest loss as the final reply.

According to the above aspect and any possible implementation manner, an implementation manner is further provided. The feature extraction in S6 is specifically: extracting text features of the final reply by using a pre-trained Bert model.

The above-mentioned aspect and any possible implementation manner further provide an implementation manner, and the classification by the model in the S6 is specifically:

The bidirectional LSTM model is used to train the corpus containing sensitive words, and the trained bidirectional LSTM classification model is used to judge whether the reply contains sensitive words; the bidirectional LSTM model is as follows:

o _t =σ(W _o [h _t - ₁ , x _t ]+b _o )

h _t =o _t *tanh(C _t )

Among them, h _t-1 represents the hidden layer state at the previous moment, x _t represents the input word at the current moment, and C _t represents the cell state at the current moment. The formula for calculating _Ct is as follows:

C _t =f _t *c _t-1 +i _t *c _t

in,

c _t =tanh(W _c ·[h _t-1 , x _t ])+b _c

f _t =σ(W _f ·[h _t-1 , x _t ]+b _f )

W is the parameter of the neural network, and b is the neuron bias.

The above-mentioned aspect and any possible implementation manner further provide a self-auditing human-machine dialogue system, the system comprising:

Input-output modules for inputting and outputting language dialogues;

The dialogue response generation module generates a candidate set composed of multiple responses according to the input language dialogue, and scores according to the MMI model, and selects the final response;

The dialogue self-audit module extracts text features according to the final reply content, classifies the text through the bidirectional LSTM classification model, and outputs machine language.

The above aspect and any possible implementation manner further provide an implementation manner, wherein the dialogue reply generation module includes:

The corpus collection unit is used to collect different types of corpus and form a corpus;

The corpus training unit trains the corpus through the GPT2 model, and splices each round of dialogues in sequence for the multiple rounds of dialogue contained in the corpus;

The encoding conversion unit converts the input language dialogue into vector encoding;

The reply sentence generation unit generates a candidate set composed of multiple replies through the corpus training unit and the encoding conversion unit;

The scoring and replying unit scores the candidate set through the MMI model, and selects the sentence with the smallest loss relative to the dialogue history in the candidate set as the final reply.

According to the above-mentioned aspect and any possible implementation manner, an implementation manner is further provided, wherein the dialogue self-audit module includes:

The feature extraction unit extracts text features for the final reply;

The model classification unit determines whether the final reply contains a sensitive word, if it contains a sensitive word, it outputs a specific sentence, and if it does not contain a sensitive word, the final reply is used as the output language dialogue.

The above aspect and any possible implementation manner further provide a computer-readable storage medium, where the computer-readable storage medium is a non-volatile storage medium or a non-transitory storage medium, in which the computer-readable storage medium is A man-machine dialogue method program is included, and when the man-machine dialogue method program is executed by the processor, the steps of a man-machine dialogue method as described in any one of the above are implemented.

Compared with the prior art, the present invention can obtain the following technical effects: the system of the present invention can realize chatting with the user, the user can input any words, and the neural network model in the dialogue reply generating module can generate multiple candidate replies, and adopt The scoring model selects replies that are more relevant to the conversation history rather than just generic replies, allowing users to have a more enjoyable chat experience. Compared with traditional generative methods, this method can improve the accuracy and relevance of chatbot responses. At the same time, the dialogue self-audit model in the dialogue system can automatically filter the replies with sensitive words, avoid affecting the user experience due to a swearing or offensive words, and always maintain a good interaction in the dialogue process.

Of course, any product implementing the present invention does not necessarily need to achieve all the above-mentioned technical effects at the same time.

【Description of drawings】

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

Fig. 1 is the work flow chart of the man-machine dialogue method of self-audit provided by one embodiment of the present invention;

FIG. 2 is a working logic diagram of a self-audit man-machine dialogue system provided by an embodiment of the present invention.

【Detailed ways】

In order to better understand the technical solutions of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

It should be understood that the described embodiments are only some, but not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. As used in the embodiments of the present invention and the appended claims, the singular forms "a," "the," and "the" are intended to include the plural forms as well, unless the context clearly dictates otherwise.

The present invention provides a self-audit man-machine dialogue system, and the self-audit man-machine dialogue method specifically includes the following steps:

S1: Collect multiple different types of corpus to form a corpus;

S2: Use the GPT2 model to train the corpus, and splicing each round of dialogue contained in the corpus in the order in which multiple rounds of dialogue are generated;

S3: Encode the input language dialogue;

S4: Generate a candidate set consisting of multiple replies through the results of S2 and S3;

S5: Score the candidate set through the MMI model, and select the sentence with the smallest loss relative to the dialogue history in the candidate set as the final reply;

S6: Perform feature extraction and model classification on the final reply, and determine whether the final reply contains sensitive words. If it contains sensitive words, output a specific sentence. If it does not contain sensitive words, use the final reply as the output language dialogue.

The S3 is specifically: converting the input language dialogue into a vector using byte pair encoding, representing a long text x ₁ , x ₂ , ·, x _n , and setting an end marker at the end.

The S4 specifically includes the following steps:

S41: Taking the input language dialogue as the source sentence, expressed as S=x ₁ , x ₂ , ·, x _m , where S is the source sentence;

S42: Use the replies in the corpus as the target sentence, expressed as T=x _m+1 , x ₂ , ·, x _n , where T is the target sentence, and the conditional probability of the target sentence is:

S43: Generate multiple candidate sentences through the autoregressive language model. The autoregressive language model uses the density estimation of the joint probability of the text sequence to model, and the formula is as follows:

S44: Select a plurality of generated candidate sentences through the Top-k Sampling method to form a candidate set composed of multiple replies.

The S5 specifically includes the following steps:

S51: splicing each candidate sentence in the candidate set with the dialogue history in reverse order;

S52: Taking each candidate sentence as the target, for the given reply hypothesis H of each candidate sentence, predict the source sentence S, and the calculation formula of the MMI model is as follows:

Among them, logp(H) represents the penalty item, and the penalty item is that the probability of the language model for the frequently occurring reply H in the training set will be higher than that of other replies, and subtracting the penalty item can reduce the score of the frequently occurring H;

S53: Select and penalize plain and general replies through the MMI model, calculate the loss of all candidate sentences relative to the dialogue history, and finally select the one with the smallest loss as the final reply.

The feature extraction in S6 is specifically: using a pre-trained Bert model to extract the text features of the final reply.

The classification by the model in the S6 is specifically:

The bidirectional LSTM model is used to train the corpus containing sensitive words, and the trained bidirectional LSTM classification model is used to judge whether the reply contains sensitive words; the bidirectional LSTM model is as follows:

o _t =σ(W _o [h _t - ₁ , x _t ]+b _o )

h _t =o _t *tanh(C _t )

Among them, h _t-1 represents the hidden layer state at the previous moment, x _t represents the input word at the current moment, and C _t represents the cell state at the current moment. The formula for calculating _Ct is as follows:

C _t =f _t *c _t-1 +i _t *c _t

in,

c _t =tanh(W _c ·[h _t-1 , x _t ])+b _c

f _t =σ(W _f ·[h _t-1 , x _t ]+b _f )

W is the parameter of the neural network, and b is the neuron bias.

The present invention also provides a self-audit man-machine dialogue system, the system includes:

Input-output modules for inputting and outputting language dialogues;

The dialogue response generation module generates a candidate set composed of multiple responses according to the input language dialogue, and scores according to the MMI model, and selects the final response;

The dialogue self-audit module extracts text features according to the final reply content, classifies the text through the bidirectional LSTM classification model, and outputs machine language.

The dialogue reply generating module includes:

The corpus collection unit is used to collect different types of corpus and form a corpus;

The corpus training unit trains the corpus through the GPT2 model, and splices each round of dialogues in sequence for the multiple rounds of dialogue contained in the corpus;

The encoding conversion unit converts the input language dialogue into vector encoding;

The reply sentence generation unit generates a candidate set composed of multiple replies through the corpus training unit and the encoding conversion unit;

The scoring and replying unit scores the candidate set through the MMI model, and selects the sentence with the smallest loss relative to the dialogue history in the candidate set as the final reply.

The dialogue self-audit module includes:

The feature extraction unit extracts text features for the final reply;

The model classification unit determines whether the final reply contains a sensitive word, if it contains a sensitive word, it outputs a specific sentence, and if it does not contain a sensitive word, the final reply is used as the output language dialogue.

The present invention also provides a computer-readable storage medium. The computer-readable storage medium is a non-volatile storage medium or a non-transitory storage medium, and the computer-readable storage medium includes a man-machine dialogue method program. When the dialogue method program is executed by the processor, the steps of any one of the man-machine dialogue methods described above are implemented.

As shown in FIG. 2 , the present invention provides a man-machine dialogue system, which is composed of an input and output module, a dialogue reply generation module, and a dialogue self-audit module. The user can enter any dialogue, generate multiple candidate responses through the generation model in the response generation module, and then calculate the loss of the candidate responses relative to the historical dialogue through the MMI scoring model, and select the one with the smallest loss as the final response. Input the generated reply into the dialogue self-audit model, first use the pre-trained Bert model to extract text features, and then use the bidirectional LSTM model for classification to determine whether it is a sensitive sentence, if not, output the reply to the user; If it is a sensitive sentence, the reply in the dialog template set in advance is output. The inventive system can be applied to any scene for chatting.

As shown in Figure 1, the specific content of the method of the present invention includes the following steps:

S1: The present invention uses corpus containing 50w multi-round dialogues and common Chinese chat corpus, such as Douban corpus, TV drama dialogue corpus, etc. to form a Chinese corpus.

S1.1: Use the GPT2 model to train the corpus, splicing the multiple rounds of dialogue contained in the Chinese corpus, and splicing each round of dialogue in sequence, for example, "[CLS] Is it better today? [SEP] It's getting worse every day. [SEP] The medicine doesn't help, go get an injection. Don't drag [SEP]".

S2: Convert the input sentence into a vector using byte pair encoding, and express it as a long text x ₁ , x ₂ , ··, x _n , and set an end marker at the end.

S3: Take the historical dialogue of a round of dialogue as the source sentence (Source), denoted as S=x ₁ , x ₂ , ·, x _m , and take the reply in the corpus as the target sentence (Target), denoted as T=x _m+1 , x ₂ , ·, x _n , the conditional probability P(T|S) is written as follows:

S4: The model of the present invention belongs to the autoregressive language model, that is, the next word is predicted according to the above content, and the natural language generation process is naturally matched. Autoregressive language models are modeled using density estimates of joint probabilities of text sequences with the following formula:

S4.1: The model of the present invention inherits the decoder of the Transformer model, and is mainly composed of 12-layer decoders. The structure of a single decoder consists of two parts, a masked multi-head self-attention mechanism and a Feed Forward Neural Network. The mask multi-head attention mechanism is the evolution of the self-attention mechanism. The self-attention mechanism is an attention that uses the dot product for similarity calculation, but only one more (the dimension of K) is used to adjust, so that the inner product will not be too large. big. Multi-head attention (Multi-head attention) mechanism, for each word in the input sentence, the attention calculation is performed with all the words in the sentence, the purpose is to learn the word dependencies inside the sentence, and capture the internal structure and semantic information of the sentence. The mask self-attention masks the information from all words to the right of the current calculation position during the self-attention calculation.

S5: In the generation part, after each token is generated, it is added to the input sequence. The new sequence will be used as the input for the next step of the model, and finally the generated sentence will be obtained. In this part, the Top-k Sampling method is used to select multiple candidate sentences to form a candidate reply set.

S6: In order to avoid generating too many bland, general and meaningless replies, the present invention adopts Mutual Information Maximization (MMI) as a scoring function to evaluate the candidate reply set.

S6.1: First, splicing each candidate reply with the dialogue history in reverse order, for example, "[CLS] taking medicine doesn't work, go for an injection. Don't drag [SEP] getting worse day by day [SEP] Today is better ? [SEP]".

S6.2: The goal is to predict the source sentence S for each given reply hypothesis H. The MMI formula is as follows:

Among them, logp(H) represents the penalty item. The meaning of this penalty item is that for the reply H that often appears in the training set, the probability of its language model will be higher than that of other replies. Subtracting this penalty item can make these frequently appearing H score dropped. Through the MMI model, you can choose to penalize plain and general replies, calculate the loss of all candidate replies relative to the dialog history, and finally select the one with the smallest loss, that is, the one closely related to the dialog history as the final reply.

S7: Send the generated reply to the dialogue self-audit module to judge whether it is a sensitive word. The self-audit model in this module consists of two parts: the extraction feature part and the classification model.

S7.1: In the feature extraction part, the pre-trained Bert model is used to extract text features.

S7.2: The classification model adopts a bidirectional LSTM model. The labeled corpus is used for training, and the training model is obtained.

The LSTM formula is as follows:

o _t =σ(W _o [h _t-1 , x _t ]+b _o )

h _t =o _t *tanh(C _t )

Among them, h _t-1 represents the hidden layer state at the previous moment, x _t represents the input word at the current moment, and C _t represents the cell state at the current moment. The formula for calculating _Ct is as follows:

C _t =f _t *c _t-1 +i _t *c _t

in,

c _t =tanh(W _c ·[h _t-1 , x _t ])+b _c

f _t =σ(W _f ·[h _t-1 , x _t ]+b _f )

W is the parameter of the neural network, and b is the neuron bias.

S7.3: Call the trained classification model to determine whether the reply contains sensitive words. If it contains sensitive words, it will output any reply in the reply template to the user, such as "Ah~ I don't know how to answer", "Hehehe~", "What you said is right!", "What the hell~ha Haha", "what are you talking about duck~", "I don't know this question yet, ask me another question~", etc.; if it is judged that no sensitive words are included, output the reply obtained in the above step S6.2 and show it to user.

The system of the present invention can realize chatting with the user, the user can input any words, multiple candidate replies can be generated through the neural network model in the dialog reply generation module, and the scoring model is used to select replies that are more relevant to the conversation history instead of just general replies , so that users have a more pleasant chat experience. Compared with traditional generative methods, this method can improve the accuracy and relevance of chatbot responses. At the same time, the dialogue self-audit model in the dialogue system can automatically filter the replies with sensitive words, avoid affecting the user experience due to a swearing or offensive words, and always maintain a good interaction in the dialogue process.

The present invention can provide free and open dialogues in multiple fields, and can complete multiple rounds of dialogues with users on a topic. This kind of dialogue without a clear task goal is more anthropomorphic and has great application value for customer marketing, education, psychotherapy and other fields. At the same time, the design of this model fully considers the dialogue history information, which is equivalent to having topic constraints in the dialogue process, avoiding other topical speeches in the process of dialogue with users, capturing the user's speech style, and always maintaining good interaction with users.

The model in the present invention adopts a self-attention mechanism. Compared with the traditional RNN or LSTM structure, it learns important information in the dialogue from multiple levels such as syntax and semantics, which endows the model with the ability to focus on different positions and avoids information overload and loss. , while avoiding excessive sequence dependencies in the model, thereby improving the parallel computing power of the model.

The dialogue self-audit module of the system of the present invention can audit the generated reply content, judge whether it is swear words or sensitive words, simulate the human thinking mode, which is equivalent to restraining the system from the moral level, and avoid sending sensitive words to users and causing bad experience , and always keep outputting appropriate and comfortable words in the dialogue scene, so that customers have a pleasant chat experience.

The self-audit man-machine dialogue method, system, and readable storage medium provided by the embodiments of the present application are described above in detail. The description of the above embodiment is only used to help understand the method of the present application and its core idea; meanwhile, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific embodiment and the scope of application, In conclusion, the content of this specification should not be construed as a limitation on the present application.

As certain terms are used in the specification and claims to refer to particular components. It should be understood by those skilled in the art that hardware manufacturers may refer to the same component by different nouns. The present specification and claims do not use the difference in name as a way to distinguish components, but use the difference in function of the components as a criterion for distinguishing. As mentioned in the entire specification and claims, "comprising" and "including" are open-ended terms, so they should be interpreted as "including/including but not limited to". "Approximately" means that within an acceptable error range, those skilled in the art can solve the technical problem within a certain error range, and basically achieve the technical effect. Subsequent descriptions in the specification are preferred embodiments for implementing the present application. However, the descriptions are for the purpose of illustrating the general principles of the present application and are not intended to limit the scope of the present application. The scope of protection of this application should be determined by the appended claims.

It should also be noted that the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion, such that a commodity or system comprising a list of elements includes not only those elements, but also includes not explicitly listed other elements, or elements inherent to the commodity or system. Without further limitation, an element defined by the phrase "comprising a..." does not preclude the presence of additional identical elements in the article or system that includes the element.

It should be understood that the term "and/or" used in this document is only an association relationship to describe the associated objects, indicating that there may be three kinds of relationships, for example, A and/or B, which may indicate that A exists alone, and A and B exist at the same time. B, there are three cases of B alone. In addition, the character "/" in this document generally indicates that the related objects are an "or" relationship.

The above description shows and describes several preferred embodiments of the present application, but as mentioned above, it should be understood that the present application is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various various other combinations, modifications and environments, and can be modified within the scope of the concept of the application described herein, using the above teachings or skill or knowledge in the relevant field. However, modifications and changes made by those skilled in the art do not depart from the spirit and scope of the present application, and should all fall within the protection scope of the appended claims of the present application.

Claims (10)

1. a man-machine dialogue method of self-examination, is characterized in that, the man-machine dialogue method of described self-examination specifically comprises the following steps: S1: Collect multiple different types of corpus to form a corpus; S2: Use the GPT2 model to train the corpus, and splicing each round of dialogue contained in the corpus in the order in which multiple rounds of dialogue are generated; S3: Encode the input language dialogue; S4: Generate a candidate set consisting of multiple replies through the results of S2 and S3; S5: Score the candidate set through the MMI model, and select the sentence with the smallest loss relative to the dialogue history in the candidate set as the final reply; S6: Perform feature extraction and model classification on the final reply, and determine whether the final reply contains sensitive words. If it contains sensitive words, output a specific sentence. If it does not contain sensitive words, use the final reply as the output language dialogue. 2. the man-machine dialogue method of self-examination according to claim 1, is characterized in that, described S3 is specifically: the language dialogue of input adopts byte pair encoding to be converted into vector, is represented as a section of long text x ₁ , x ₂ , ·, x _n , and set the closing tag at the end. 3. The method according to claim 2, wherein the S4 specifically comprises the following steps: S41: Take the input language dialogue as the source sentence, and express it as S=x ₁ , x ₂ , ·, x _m , where S is the source sentence; S42: Use the replies in the corpus as the target sentence, expressed as T=x _m+1 , x ₂ , ·, x _n , where T is the target sentence, and the conditional probability of the target sentence is:

S43: Generate multiple candidate sentences through the autoregressive language model. The autoregressive language model uses the density estimation of the joint probability of the text sequence to model, and the formula is as follows:

S44: Select a plurality of generated candidate sentences through the Top-k Sampling method to form a candidate set composed of multiple replies. 4. The man-machine dialogue method of self-audit according to claim 3, is characterized in that, described S5 specifically comprises the following steps: S51: splicing each candidate sentence in the candidate set with the dialogue history in reverse order; S52: Taking each candidate sentence as the target, for the given reply hypothesis H of each candidate sentence, predict the source sentence S, and the calculation formula of the MMI model is as follows:

Among them, logp(H) represents the penalty item, and the penalty item is that the probability of the language model for the frequently occurring reply H in the training set will be higher than that of other replies, and subtracting the penalty item can reduce the score of the frequently occurring H; S53: Select and penalize plain and general replies through the MMI model, calculate the loss of all candidate sentences relative to the dialogue history, and finally select the one with the smallest loss as the final reply. 5 . The method according to claim 1 , wherein the feature extraction in the S6 is specifically: using a pre-trained Bert model to extract the text features of the final reply. 6 . 6. The man-machine dialogue method of self-examination according to claim 3, is characterized in that, classifying by model in described S6 is specifically: The bidirectional LSTM model is used to train the corpus containing sensitive words, and the trained bidirectional LSTM classification model is used to judge whether the reply contains sensitive words; the bidirectional LSTM model is as follows: o _t =σ(W _o [h _t-1 , x _t ]+b _o ) h _t =o _t *tanh(C _t ) Among them, h _t-1 represents the hidden layer state at the previous moment, x _t represents the input word at the current moment, C _t represents the cell state at the current moment, and the calculation formula of C _t is as follows: C _t =f _t *c _t-1 +i _t *c _t in, c _t =tanh(W _c ·[h _t-1 , x _t ])+b _c f _t =σ(W _f ·[h _t-1 , x _t ]+b _f ) W is the parameter of the neural network, and b is the neuron bias. 7. A self-audited man-machine dialogue system, comprising the self-audited man-machine dialogue method of one of the above claims 1-6, wherein the system comprises: Input-output module for inputting and outputting language dialogues; The dialogue response generation module generates a candidate set composed of multiple responses according to the input language dialogue, and scores according to the MMI model, and selects the final response; The dialogue self-audit module extracts text features according to the final reply content, classifies the text through the bidirectional LSTM classification model, and outputs machine language. 8. The man-machine dialogue system for self-examination according to claim 7, wherein the dialogue reply generating module comprises: The corpus collection unit is used to collect different types of corpus and form a corpus; The corpus training unit trains the corpus through the GPT2 model, and splices each round of dialogues in sequence for the multiple rounds of dialogue contained in the corpus; The encoding conversion unit converts the input language dialogue into vector encoding; The reply sentence generation unit generates a candidate set composed of multiple replies through the corpus training unit and the code conversion unit; The scoring and replying unit scores the candidate set through the MMI model, and selects the sentence with the smallest loss relative to the dialogue history in the candidate set as the final reply. 9. The man-machine dialogue system of self-examination according to claim 8, wherein the dialogue self-examination module comprises: The feature extraction unit extracts text features for the final reply; The model classification unit determines whether the final reply contains a sensitive word, if it contains a sensitive word, it outputs a specific sentence, and if it does not contain a sensitive word, the final reply is used as the output language dialogue. 10. A computer-readable storage medium, the computer-readable storage medium is a non-volatile storage medium or a non-transitory storage medium, and the computer-readable storage medium comprises a man-machine dialogue method program, the man-machine dialogue method When the program is executed by the processor, the steps of the method for man-machine dialogue according to any one of claims 1 to 6 are implemented.

Images