CN114547274B - Multi-turn question and answer method, device and equipment - Google Patents

Abstract

The application provides a method, a device and equipment for multi-turn question answering. According to the method, for a new round of initiated conversation, the current round of conversation information and historical conversation information are input into a context rewriting model, and the trained context rewriting model is used for determining an object to be rewritten and a corresponding rewriting operation type in the current round of conversation information and a key text for rewriting in the historical conversation information; rewriting the current round of dialogue information by using a key text according to the objects to be rewritten in the current round of dialogue information and the rewriting operation types corresponding to the objects to be rewritten, and generating candidate dialogue information; and one candidate dialogue information with better quality is used as the rewritten target dialogue information, so that the omitted key information in the current round of dialogue is completely supplemented, the designated information is replaced by the designated complete content information, the problems of designation and omission in multiple rounds of questions and answers are solved, the corresponding reply information is obtained based on the rewritten target dialogue information, and the quality and the accuracy of the reply information are improved.

Description

Method, device and equipment for multi-round question and answer

technical field

The present application relates to the field of computer technology, and in particular, to a method, apparatus and device for multi-round question and answer.

Background technique

The human-computer interaction system (Table QA) based on tabular data is to convert the natural language questions input by users into corresponding Structured Query Language (SQL) statements, and store tabular data in the database. The function of finding the result corresponding to the SQL statement in the table and feeding it back to the user. Since tables are a common data storage structure in daily work and life, Table QA system has a wide range of application scenarios, such as data query, statistics and screening, etc., and can be applied to various application fields such as government affairs, finance, and energy. The key technology to realize the Table QA system is to convert natural language questions into corresponding SQL statements, that is, Text-to-SQL language conversion.

In practical application scenarios, users often interact with the system in the form of multiple rounds of Q&A. There are often redundant content in the historical multiple rounds of Q&A. This round of dialogue may omit or use other content to refer to the historical dialogue. message appeared. In order to enable the dialogue system to have the ability to remember, some important information in the dialogue history is usually stored through dialogue state tracking (Dialog State Tracking, DST for short), but DST-related methods are often complex and cannot be applied to online human-computer interaction. system.

In the current human-computer interaction system, the text-to-SQL language conversion model is input by splicing all the historical dialogue content of the current round and the historical rounds, which introduces additional noise, and the converted SQL statement is inaccurate, so As a result, the response information given by the human-computer interaction system is of poor quality and low accuracy.

SUMMARY OF THE INVENTION

The present application provides a method, device and device for multi-round question and answer, which are used to solve the problems of poor quality and low accuracy of answer information given by the current human-computer interaction system.

In a first aspect, the present application provides a method for multiple rounds of question and answer, including:

In response to the dialogue request, if it is determined that there is at least one round of historical dialogue before the current round of dialogue, acquiring the current round of dialogue information and the historical dialogue information of the at least one round of historical dialogue;

Input the current round of dialogue information and the historical dialogue information into the context rewriting model, and use the context rewriting model to determine the object to be rewritten in the current round of dialogue information, the type of rewriting operation corresponding to the object to be rewritten, and the The key text used for rewriting in the historical dialogue information;

According to the object to be rewritten in the current round of dialogue information and the rewriting operation type corresponding to the object to be rewritten, use the key text to rewrite the current round of dialogue information to generate candidate dialogue information;

According to the quality information of the candidate dialogue information, one of the candidate dialogue information is used as the target dialogue information after the current round of dialogue information is rewritten;

Acquire reply information corresponding to the target dialogue information, and feed back the reply information.

In a second aspect, the present application provides a multi-round question and answer device, comprising:

an information acquisition module, configured to, in response to the dialogue request, acquire the dialogue information of the current round and the historical dialogue information of the at least one round of historical dialogue if it is determined that there is at least one round of historical dialogue before the current round of dialogue;

The rewriting prediction module is used to input the current round of dialogue information and the historical dialogue information into a context rewriting model, and use the context rewriting model to determine the object to be rewritten in the current round of dialogue information, and the corresponding object to be rewritten. The type of rewriting operation, and the key text used for rewriting in the historical dialogue information;

A rewriting module, configured to rewrite the current round of dialogue information by using the key text according to the object to be rewritten and the rewriting operation type corresponding to the object to be rewritten in the current round of dialogue information to generate candidate dialogue information;

a selection module, configured to use one of the candidate dialogue information as the target dialogue information after rewriting the current round of dialogue information according to the quality information of the candidate dialogue information;

A reply module, configured to obtain reply information corresponding to the target dialogue information, and feed back the reply information.

In a third aspect, the present application provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor;

the memory stores computer-executable instructions;

The processor executes the computer-executable instructions stored in the memory to implement the method described in the first aspect above.

In a fourth aspect, the present application provides a computer-readable storage medium, where computer-executable instructions are stored in the computer-readable storage medium, and the computer-executable instructions are used to implement the method described in the first aspect when executed by a processor. .

In a fifth aspect, the present application provides a computer program product, including a computer program, which implements the method described in the first aspect when the computer program is executed by a processor.

The method, device and device for multi-round question and answer provided by the present application, for a new round of dialogue (this round of dialogue) initiated, by obtaining the historical dialogue information of at least one round of historical dialogues before the current round of dialogue, the current round of dialogue information and The historical dialogue information is input into the context rewriting model, and the trained context rewriting model is used to determine the object to be rewritten in this round of dialogue information, the type of rewriting operation corresponding to the object to be rewritten, and the key text used for rewriting in the historical dialogue information; In the dialogue information, the object to be rewritten and the type of rewriting operation corresponding to the object to be rewritten, use the key text to rewrite the current round of dialogue information to generate candidate dialogue information; according to the quality information of the candidate dialogue information, a candidate dialogue information with better quality As the target dialogue information after the current round of dialogue information is rewritten, it can complete the key information omitted in the current round of dialogue, and replace the reference information with the complete content information referred to, so as to rewrite the current round of dialogue information to high quality It solves the problems specified and omitted in multiple rounds of question and answer, and obtains the corresponding reply information based on the rewritten target dialogue information, which improves the quality and accuracy of the reply information.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description serve to explain the principles of the application.

Fig. 1 is an example diagram of a plurality of rounds of question and answer content provided by this application;

2 is a schematic diagram of an exemplary network architecture on which this application is based;

3 is a flowchart of a method for a multi-round question and answer provided by an exemplary embodiment of the present application;

4 is a flowchart of a method for multiple rounds of question and answer provided by another exemplary embodiment of the present application;

FIG. 5 is a framework diagram used by a context rewriting model provided by an exemplary embodiment of the present application;

6 is a framework diagram of a Text-to-SQL language conversion based on a context rewriting model provided by an exemplary embodiment;

7 is a schematic structural diagram of an apparatus for multiple rounds of question and answer provided by an exemplary embodiment of the present application;

FIG. 8 is a schematic structural diagram of an electronic device provided by an exemplary embodiment of the present application.

Specific embodiments of the present application have been shown by the above-mentioned drawings, and will be described in more detail hereinafter. These drawings and written descriptions are not intended to limit the scope of the concepts of the present application in any way, but to illustrate the concepts of the present application to those skilled in the art by referring to specific embodiments.

Detailed ways

Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. Where the following description refers to the drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the illustrative examples below are not intended to represent all implementations consistent with this application. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present application as recited in the appended claims.

First, the terms involved in this application are explained:

Data table: The data table in the Structured Query Language (SQL) database is used to store data. After the data is stored in the corresponding data table of the database, the permanentization of the data to be stored is completed, and then the data table of the database can be queried by accessing , to get the data content of the data.

Data element: also known as the schema of the data table, refers to the table name, column name, value and other information in the data table.

Table QA: A human-computer interaction system based on table data (or table knowledge), which converts natural language questions input by users into SQL statements, and finds the corresponding SQL statements in the data table used to store table data in the database The function of feeding back the results to the user. For example, a question answering/dialogue system based on tabular data.

Text-to-SQL language conversion model: It is the core technology of the Table QA system, a method of semantic parsing, which is used to convert questions (Text) described by human natural language into computer-executable SQL statements. With this model, free interaction between people and tables/databases can be achieved without requiring users to learn complex SQL syntax. Further, a table/database-based question answering/dialogue system can be implemented by relying on this model.

token: Each word (or word) obtained after the text is tokenized in natural language processing.

In addition, the terms "first", "second", etc. are used for descriptive purposes only, and should not be construed as indicating or implying relative importance or implying the number of indicated technical features. In the description of the following embodiments, the meaning of "plurality" is two or more, unless otherwise expressly and specifically defined.

The multi-round question answering method provided in this application can be applied to human-computer interaction systems in various application fields, for example, question answering/dialogue systems based on tabular data (tabular knowledge or database).

In the actual human-computer interaction system, users often interact with the system in the form of multiple rounds of question and answer. There are often redundant content in the historical multiple rounds of question and answer content. This round of dialogue may omit or use other content to refer to the historical dialogue. Information has already appeared in .

Exemplarily, taking the multi-round question-and-answer content shown in FIG. 1 as an example, in this example, “they” in the current (third round) dialogue refers to the players queried based on the two rounds of historical dialogues, that is, "Poker players over 200cm tall", the complete question in this round of dialogue can be expressed as "What is the average height of poker players over 200cm tall".

In the traditional human-computer interaction system, the content of the above three rounds of dialogues are all spliced and input into the Text-to-SQL language conversion model, which introduces a lot of extra noise, and the converted SQL statements are inaccurate, resulting in the error given by the human-computer interaction system. The response information is of poor quality and low accuracy.

In view of the above technical problems, the present application provides a method for multiple rounds of question and answer, inputting the current round of dialogue information and historical dialogue information into a context rewriting model, and using the trained context rewriting model to determine the object to be rewritten in the current round of dialogue information, to be rewritten The type of rewriting operation corresponding to the object, and the key text used for rewriting in the historical dialogue information; according to the type of rewriting operation corresponding to the object to be rewritten and the object to be rewritten in the current round of dialogue information, use the key text to rewrite the current round of dialogue information, Generate candidate dialogue information; according to the quality information of the candidate dialogue information, a candidate dialogue information with better quality is used as the target dialogue information after the current round of dialogue information rewrites, which can supplement the key information omitted in the current round of dialogue. The substitute information is replaced with the complete content information referred to, so that the current round of dialogue information is rewritten into a high-quality expression, the problem of specifying and omitting in multiple rounds is solved, and the corresponding reply information is obtained based on the rewritten target dialogue information, which improves the Quality and accuracy of response information.

Exemplarily, FIG. 2 is a schematic diagram of an exemplary network architecture on which this application is based, and the network architecture shown in FIG. 2 may specifically include a server and a terminal.

Specifically, the server may be a server cluster set up in the cloud, and the server stores historical dialogue data and data stored in the database for realizing human-computer interaction data query, such as table data for realizing form question answering/dialogue, etc. The server also stores related model data such as the context rewriting model, the Text-to-SQL language conversion model, etc. Through the preset computing logic in the server, the server can realize various computing functions such as model training and multi-round question-and-answer human-computer interaction. .

Specifically, the terminal may be a hardware device with network communication function, computing function, and information display function, which includes but is not limited to a smart phone, a tablet computer, a desktop computer, an Internet of Things device, and the like.

Through the communication interaction with the server, the terminal sends the current round of dialogue information input by the user to the server. After the server obtains the current dialogue information input by the user, if there is at least one historical dialogue of the current dialogue before the current dialogue, the server acquires the historical dialogue information of at least one historical dialogue, and rewrites the model through the trained context, according to The historical dialogue information rewrites the current round of dialogue information to obtain the rewritten target dialogue information; input the target dialogue information into the Text-to-SQL language conversion model to convert the target dialogue information into the corresponding SQL statement; query data according to the SQL statement The form gets the reply information, and feeds back the reply information to the terminal.

The technical solutions of the present application and how the technical solutions of the present application solve the above-mentioned technical problems will be described in detail below with specific examples. The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

FIG. 3 is a flowchart of a method for multiple rounds of question and answer provided by an exemplary embodiment of the present application. The method for multiple rounds of question and answer provided in this embodiment can be specifically applied to the aforementioned server. As shown in Figure 3, the specific steps of the method are as follows:

Step S301 , in response to the dialogue request, if it is determined that there is at least one round of historical dialogue before the current round of dialogue, acquire the current round of dialogue information and the historical dialogue information of at least one round of historical dialogues.

In practical applications, the user can input the content of the dialogue on the interactive page provided on the terminal, and the terminal will display the reply information fed back by the service to the user through the interactive page. After the user inputs a new round of dialogue content on the terminal, the terminal sends a dialogue request to the server, the dialogue request includes the current dialogue text information input by the user, and is used to request the server to give reply information according to the current dialogue text information.

The server extracts the text information of the current round of dialogues in response to the dialogue request, and determines whether there are one or more rounds of historical dialogues before the current round of dialogues based on the stored historical dialogue data. If it is determined that there is at least one round of historical dialogue before the current round of dialogue, historical dialogue information of at least one round of historical dialogue is acquired.

Exemplarily, the current round of dialogue information obtained by the server may be the word segmentation result corresponding to the current round of dialogue text information, that is, a token sequence obtained after word segmentation of the current round of dialogue text.

Exemplarily, in this embodiment, word segmentation for Chinese text is a sequence of converting a sentence into a single character (eg, a single Chinese character, a single number, and a single letter). Tokenization for English text is the transformation of a sentence into a sequence of individual words (English words).

For example, for "what is their average height", the token sequence after tokenizing the sentence is {he, they, of, average, average, height, height, yes, more, less}.

For example, for "how many players are taller than 200cm", the token sequence after tokenizing the sentence is {yes, more, less, play, home, at, height, over, over, 2, 0, 0 , c, m }

Step S302, input the current round of dialogue information and historical dialogue information into the context rewriting model, and use the context rewriting model to determine the object to be rewritten in the current round of dialogue information, the rewriting operation type corresponding to the object to be rewritten, and the historical dialogue information for rewriting. key text.

After obtaining the current round of dialogue information and the historical dialogue information of one or more rounds of historical dialogues in this interaction, input the current round of dialogue information and historical dialogue information into the trained context rewriting model, and use the context rewriting model to Determine whether the text information at each position in the current round of dialogue information is retained, replaced or deleted, and at which position the key information in the historical dialogue needs to be inserted, and extracting from the historical dialogue information may be added to the current round of dialogue information The key information is used as the key text to rewrite the dialogue information of this round.

In this embodiment, the context rewriting model is a pre-trained neural network model, which is used to determine the rewriting position category and rewriting operation type corresponding to each position in the current round of dialogue information according to the inputted current round of dialogue information and historical dialogue information, and the candidate position category corresponding to each position in the historical dialogue information, so as to further determine the object to be rewritten in the current round of dialogue information, the type of rewriting operation corresponding to the object to be rewritten, and the key text used for rewriting in the historical dialogue information.

The object to be rewritten may be the position to be inserted or the text to be rewritten.

The position to be inserted is the corresponding position of any token in the current round of dialogue information, and each token in the current round of dialogue information corresponds to a position. The key information in the historical dialogue needs to be inserted before the to-be-inserted position.

The current round of dialogue information may include one or more texts to be rewritten, and one to-be-rewritten text includes one or more consecutive tokens in the current round of dialogue information. The key text used for rewriting in the historical dialogue information may be one or more key texts, and each key text includes one or more consecutive tokens in the historical dialogue information.

Illustratively, the type of overwrite operation includes at least one of the following: insert, delete, replace, keep as is.

The type of rewriting operation corresponding to the position to be inserted is insert, and the type of rewriting operation corresponding to the text to be rewritten may be any one of deletion, replacement, and keeping as it is.

For the target position where the rewrite operation type is insertion, each target position to be inserted is processed separately, and a candidate token can be inserted into this position.

Step S303 , according to the object to be rewritten and the rewriting operation type corresponding to the object to be rewritten in the dialogue information of the current round, use the key text to rewrite the dialogue information of the current round to generate candidate dialogue information.

After determining the object to be rewritten in the current round of dialogue information and the type of rewriting operation corresponding to the object to be rewritten, and the key text used for rewriting in the historical dialogue information, the current round of dialogue information is automatically rewritten to generate rewritten candidate dialogue information.

Exemplarily, the current round of dialogue information may include one or more texts to be rewritten, and the key texts used for rewriting in the historical dialogue information may be one or more key texts, which can be used according to the type of rewriting operation corresponding to the object to be rewritten. The enumeration method gives all possible rewriting methods and generates all possible candidate dialogue information.

In an actual interactive system, the content of the dialogue is usually relatively simple, and the current round of dialogue information usually contains a text to be rewritten whose type of rewriting operation is replacement, or a position to be inserted.

Exemplarily, if the current round of dialogue information includes only one position to be inserted, the sequence of all key texts can be arranged and combined, all possible arrangements can be enumerated, the corresponding combined texts can be obtained, and each combined text can be inserted into the current round. The to-be-inserted position in the dialogue information is obtained to obtain the corresponding candidate dialogue information.

Exemplarily, if the current round of dialogue information includes only one text to be rewritten, and the rewriting operation type of the text to be rewritten is replacement, the order of all key texts can be arranged and combined, and all possible arrangements can be enumerated to obtain the corresponding Combining text, replacing the text to be rewritten in the current round of dialogue information with each combined text to obtain corresponding candidate dialogue information.

For example, taking the multi-round question and answer content shown in FIG. 1 as an example, using the context rewriting model, it can be determined that the object to be rewritten in this round of dialogue information is "they", and the corresponding rewriting operation type is: replace. Key text in the historical dialogue information includes: "Over 200cm tall" and "Poker player". Through enumeration, at least the following candidate dialogue information can be generated: "What is the average height of poker players over 200cm tall", "What is the average height of poker players over 200cm tall".

Exemplarily, if the current round of dialogue information includes multiple objects to be rewritten that need to use key text (including the position to be inserted and the text to be rewritten whose type of rewriting operation is replacement), then each to-be-rewritten corresponds to at least one key text, through Enumerate all possible rewriting methods, and generate candidate dialog information corresponding to each rewriting method.

Step S304 , according to the quality information of the candidate dialogue information, use one of the candidate dialogue information as the target dialogue information after the current round of dialogue information is rewritten.

After rewriting the dialogue information of the current round to obtain candidate dialogue information, if there are multiple candidate dialogue information, it can be sorted according to the quality information of the candidate dialogue information, and a candidate dialogue information with better quality is used as the rewritten dialogue information of the current round. Target dialogue information.

In addition, if there is only one candidate dialogue information, the candidate dialogue information is directly used as the target dialogue information after the current round of dialogue information is rewritten.

The target dialogue information after the current round of dialogue information is rewritten, completes the key information omitted in the current round of dialogue, and replaces the reference information with the complete content information referred to, which is a high-quality expression text.

Step S305: Acquire reply information corresponding to the target dialogue information, and feed back the reply information.

After determining the target dialog information after the rewritten dialog information of the current round is determined, corresponding reply information is generated based on the target dialog information, and a high-quality reply to the current dialog of the user can be obtained. The server feeds back the obtained reply information to the terminal.

In this embodiment, for a new round of dialogue (this round of dialogue) initiated, obtain the historical dialogue information of at least one round of historical dialogues before the current round of dialogue, input the current round of dialogue information and historical dialogue information into the context rewriting model, and use the training A good context rewriting model determines the object to be rewritten in this round of dialogue information, the type of rewriting operation corresponding to the object to be rewritten, and the key text used for rewriting in historical dialogue information; according to the object to be rewritten and the object to be rewritten in this round of dialogue information For the corresponding rewriting operation type, use the key text to rewrite the current round of dialogue information to generate candidate dialogue information; according to the quality information of the candidate dialogue information, a candidate dialogue information with better quality is used as the target dialogue after the current round of dialogue information is rewritten. It can supplement the key information omitted in the current round of dialogue, and replace the reference information with the complete content information referred to, so as to rewrite the current round of dialogue information into high-quality expressions and solve the problems specified in multiple rounds of question and answer. For the omitted questions, the corresponding reply information is obtained based on the rewritten target dialogue information, which improves the quality and accuracy of the reply information.

In an optional embodiment, the context rewriting model is used to determine whether the text information at each position in the current round of dialogue information is retained, replaced or deleted, and before which position the key information in the historical dialogue needs to be inserted, and from The key information that may be added to the current round of dialogue information is extracted from the historical dialogue information as key texts to rewrite the current round of dialogue information.

Specifically, input the current round of dialogue information and historical dialogue information into the context rewriting model, and perform classification processing through the context rewriting model to obtain a classification result containing the following information:

The rewriting position category and the rewriting operation type corresponding to each position in the current dialogue information, the candidate position category corresponding to each position in the historical dialogue information, wherein the rewriting position category is the rewriting starting position, the rewriting ending position, and the non-rewriting starting and ending positions. Any of ; the candidate position category is any one of the candidate start position, the candidate end position, and the non-candidate start and end position.

Further, according to the classification result, the object to be rewritten in the current round of dialogue information, the rewriting operation type corresponding to the object to be rewritten, and the key text in the historical dialogue information are determined.

Referring to FIG. 4 , FIG. 4 is a flowchart of a method for multiple rounds of question and answer provided by another exemplary embodiment of the present application. As shown in Figure 4, the specific steps of the method are as follows:

Step S401 , in response to the dialogue request, if it is determined that there is at least one round of historical dialogue before the current round of dialogue, acquire the current round of dialogue information and the historical dialogue information of at least one round of historical dialogues.

In practical applications, the user can input the content of the dialogue on the interactive page provided on the terminal, and the terminal will display the reply information fed back by the service to the user through the interactive page. After the user inputs a new round of dialogue content on the terminal, the terminal sends a dialogue request to the server, the dialogue request includes the current dialogue text information input by the user, and is used to request the server to give reply information according to the current dialogue text information.

The server extracts the text information of the current round of dialogues in response to the dialogue request, and determines whether there are one or more rounds of historical dialogues before the current round of dialogues based on the stored historical dialogue data. If it is determined that there is at least one round of historical dialogue before the current round of dialogue, historical dialogue information of at least one round of historical dialogue is acquired.

Exemplarily, the current round of dialogue information obtained by the server may be the word segmentation result corresponding to the current round of dialogue text information, that is, a token sequence obtained after word segmentation of the current round of dialogue text.

Optionally, in this step, the current round of dialogue texts and the historical dialogue texts of at least one round of historical dialogues are obtained; the current round of dialogue texts and the historical dialogue texts of each round of historical dialogues are subjected to word segmentation respectively to obtain the token of the current round of dialogue texts. Sequence and token sequence of historical dialogue text for each round of historical dialogue.

Exemplarily, in this embodiment, word segmentation for Chinese text is a sequence of converting a sentence into a single character (eg, a single Chinese character, a single number, and a single letter). Tokenization for English text is the transformation of a sentence into a sequence of individual words (English words).

For example, for "what is their average height", the token sequence after tokenizing the sentence is {he, they, of, average, average, height, height, yes, more, less}.

For example, for "how many players are taller than 200cm", the token sequence after tokenizing the sentence is {yes, more, less, play, home, at, height, over, over, 2, 0, 0 , c, m }

Through steps S402-S404, based on the context rewriting model, the target token to be rewritten in the dialogue information of the current round, the rewriting operation type corresponding to the target token, and the candidate token for rewriting in the historical dialogue information are determined.

Step S402, input the current round of dialogue information and historical dialogue information into the context rewriting model, perform classification processing through the context rewriting model, and obtain a classification result containing the following information: the rewriting location category and the rewriting operation type corresponding to each position in the current round of dialogue information , the candidate position category corresponding to each position in the historical dialogue information.

Here, the rewriting position type is any one of a rewriting start position, a rewriting end position, and a non-rewriting start and end position.

The candidate position category is any one of a candidate start position, a candidate end position, and a non-candidate start and end position.

In this embodiment, the context rewriting model is a pre-trained neural network model, which is used to determine whether the text information at each position in the current round of dialogue information is to be retained, replaced or deleted according to the inputted current round of dialogue information and historical dialogue information , and before which position to insert the key information in the historical dialogue, and extract the key information that may be added to the current round of dialogue information from the historical dialogue information as the key text to rewrite the current round of dialogue information, so as to determine The rewrite location category and rewrite operation type corresponding to each position in the current dialogue information, and the candidate position category corresponding to each position in the historical dialogue information.

Optionally, the context rewriting model may include a backbone network and a classifier, where the backbone network is a Transformer model, specifically, BART or T5. The classifiers include the following three classifiers: a position prediction classifier, a reference prediction classifier, and a replacement prediction classifier. Input the current round of dialogue information and historical dialogue information into the backbone network of the context rewriting model for processing, and input the processing results into three classifiers for classification processing to obtain the classification results.

Wherein, the location prediction classifier is used to determine the rewrite location category corresponding to each location in the current round of dialogue information, and the predictive classifier is used to determine the rewrite operation type corresponding to each location in the current round of dialogue information. Based on the classification results of the two classifiers, it can be determined whether the text information at each position in the current round of dialogue information is retained, replaced, deleted, or inserted before the position.

The replacement prediction classifier is used to determine the candidate position category corresponding to each position in the historical dialogue information. According to the classification result of the classifier, the key information that may be added to the current round of dialogue information can be extracted from the historical dialogue information as the key text.

Exemplarily, the classifier in the context rewriting model may be implemented by using a multi-layer perceptron (Multilayer Perceptron, MLP for short).

Exemplarily, as shown in FIG. 5 , the context rewriting model may add three MLPs on the basis of the Transformer, and each MLP is used as a classifier to determine the corresponding classification result. Based on the multi-round question and answer content shown in Figure 1, as shown in Figure 5, the dialogue information of the current round of dialogue and two rounds of historical dialogue is input into the Transformer, and the processing results of the Transformer are input into three classifiers to output the classification shown in Figure 5. result.

In this embodiment, the context rewriting model is a model trained based on the training set.

Exemplarily, before performing model training, first obtain a large number of historical data samples, including: current round of dialogue samples, at least one round of historical dialogue samples and target dialogue samples after rewriting the current round of dialogue information. By analyzing the current round of dialogue samples, at least one round of historical dialogue samples and target dialogue samples, it is possible to determine the rewrite location category and rewrite operation type corresponding to each position in this round of dialogue samples, as well as the corresponding rewrite operation type for each position in the historical dialogue samples. The candidate position category is obtained, and the labeled data is obtained. In addition, the annotation data can also be obtained by manual annotation.

The training set includes multiple pieces of training data, each piece of training data includes the current round of dialogue samples, at least one round of historical dialogue samples, and labeled data (including the rewrite location category and rewrite operation type corresponding to each position in the current round of dialogue samples, and historical the candidate position category corresponding to each position in the dialogue sample).

During model training, input the current round of dialogue samples and at least one round of historical dialogue samples into the context rewriting model to obtain the classification result of the context rewriting model, calculate the loss according to the classification result of the context rewriting model and the labeled data, and optimize the context rewriting according to the loss Model parameters for the model. After multiple rounds of iterative training to the model convergence, determine the trained context to rewrite the model.

Optionally, during model training, the prediction result of the model on the test set may be rewritten based on the context, and the rewritten target dialogue information may be determined. Use the perplexity degree (or sentence probability) of the rewritten target dialogue information to measure the pros and cons of the context rewriting model. The lower the perplexity degree (or the higher the sentence probability), the better the context rewriting model, which can make use of the trained context rewriting model. The rewritten target dialogue information determined by the model is of higher quality and smoother.

Step S403 , according to the classification result, determine the object to be rewritten in the dialogue information of the current round, the rewriting operation type corresponding to the object to be rewritten, and the key text in the historical dialogue information.

The object to be rewritten is the position to be inserted or the text to be rewritten. Types of overwrite operations include: insert, delete, replace, and keep-as-is. The rewrite operation type corresponding to the position to be inserted is insert.

After the classification result is obtained, an object to be rewritten is determined according to each rewriting start position and the corresponding rewriting termination position in the classification result.

The rewriting operation type of the object to be rewritten is determined according to the rewriting operation type of the location of each object to be rewritten.

Exemplarily, in the classification result, the rewriting start position and the rewriting termination position appear in pairs, and one rewriting start position corresponds to the rewriting termination position that occurs first afterward. The overwriting start position and the overwriting end position appearing in pairs determine an object to be overwritten, which includes the text at the overwriting start position, the overwriting end position, and the positions in between. The rewriting start position and the rewriting end position that appear in pairs, and the rewriting operation types of the positions between the two are consistent, that is, the rewriting operation types of the corresponding objects to be rewritten.

Exemplarily, if the rewriting start position and the rewriting termination position coincide, it is determined that the object to be rewritten corresponds to one position. If the rewriting operation type of the position is any one of replace, delete, and keep as it is, the object to be rewritten is the text to be rewritten, and the operation type corresponding to the text to be rewritten is consistent with the rewriting operation type of the position. If the rewriting operation type of the location is insert, the object to be rewritten is the location to be inserted, and the operation type corresponding to the location to be inserted is insert. Subsequent key text is inserted before (or after) the position to be inserted.

For example, according to the classification result shown in Figure 5: the rewriting start position in the current round of dialogue information is 0, and the rewriting end position is 1, and it is determined that the object to be rewritten is the text "they" on the position 0-1, which is to be rewritten. The action type is Replace.

In this step, a key text is determined according to each candidate start position and corresponding candidate end position in the classification result.

Exemplarily, in the classification result, the candidate start position and the candidate end position appear in pairs, and one candidate start position corresponds to the candidate end position that appears first afterward. The paired candidate start position and candidate end position determine a key text that includes the candidate start position, the candidate end position, and the text at all of these positions in between.

For example, according to the classification result shown in Fig. 5, it can be determined that the candidate starting position in the historical dialogue information of round 1 is 6, the candidate ending position is 9, and the key text is determined to be the text on position [6-9] "Poker player"; the candidate starting position in the historical dialogue information of the second round is 6, the candidate ending position is 14, and the key text is determined to be the text "height over 200cm" on position [6-14].

Step S404 , according to the object to be rewritten and the rewriting operation type corresponding to the object to be rewritten in the dialogue information of the current round, use the key text to rewrite the dialogue information of the current round to generate candidate dialogue information.

After determining the object to be rewritten in the current round of dialogue information and the type of rewriting operation corresponding to the object to be rewritten, and the key text used for rewriting in the historical dialogue information, the current round of dialogue information is automatically rewritten to generate rewritten candidate dialogue information.

Specifically, if there is a position to be inserted in the current round of dialogue information, at least one key text is inserted into the position to be inserted. It should be noted that the to-be-inserted position indicates that key text is to be inserted before (or after) the to-be-inserted position.

If there is the first text to be rewritten in the current round of dialogue information, and the rewriting operation type of the first text to be rewritten is replacement, at least one key text is used to replace the first text to be rewritten.

If the second to-be-rewritten text exists in the current round of dialogue information, and the rewriting operation type of the second to-be-rewritten text is delete, the second to-be-rewritten text in the current round of dialogue information is deleted.

In an actual interactive system, the content of the dialogue is usually relatively simple, and the current round of dialogue information usually contains a text to be rewritten whose type of rewriting operation is replacement, or a position to be inserted.

Exemplarily, if the current round of dialogue information includes only one position to be inserted, the sequence of all key texts can be arranged and combined, all possible arrangements can be enumerated, the corresponding combined texts can be obtained, and each combined text can be inserted into the current round. The to-be-inserted position in the dialogue information is obtained to obtain the corresponding candidate dialogue information.

Exemplarily, if the current round of dialogue information includes only one text to be rewritten, and the rewriting operation type of the text to be rewritten is replacement, the order of all key texts can be arranged and combined, and all possible arrangements can be enumerated to obtain the corresponding Combining text, replacing the text to be rewritten in the current round of dialogue information with each combined text to obtain corresponding candidate dialogue information.

Exemplarily, if the current round of dialogue information includes multiple objects to be rewritten that need to use key text (including the position to be inserted and the text to be rewritten whose type of rewriting operation is replacement), then each to-be-rewritten corresponds to at least one key text, through Enumerate all possible rewriting methods, and generate candidate dialog information corresponding to each rewriting method.

For example, taking the classification result shown in Fig. 5 as an example, the determined text to be rewritten in the current round of dialogue information is the text "they" at position 0-1, and the operation type to be rewritten is replacement; the key text in the historical dialogue information Including: the text "Poker Player" on position [6-9] in Round 1, and the text "Over 200cm tall" on Position [6-14] in Round 2. Through enumeration, at least the following candidate dialogue information can be generated: "What is the average height of poker players over 200cm tall", "What is the average height of poker players over 200cm tall".

Step S405: According to the quality information of the candidate dialogue information, use one of the candidate dialogue information as the target dialogue information after the rewriting of the dialogue information of the current round.

After rewriting the dialogue information of the current round to obtain candidate dialogue information, if there are multiple candidate dialogue information, it can be sorted according to the quality information of the candidate dialogue information, and a candidate dialogue information with better quality is used as the rewritten dialogue information of the current round. Target dialogue information.

Optionally, the sentence probability of each candidate dialogue information is obtained, and the candidate dialogue information with the highest sentence probability is used as the target dialogue information.

Exemplarily, the sentence probability of each candidate dialogue information can be calculated based on the n-gram model, or it can be implemented by any existing method of calculating the sentence probability, which is not specifically limited here.

Optionally, the perplexity degree of each candidate dialogue information is obtained, and the candidate dialogue information with the lowest perplexity degree is used as the target dialogue information.

Among them, the perplexity of a sentence W can be determined as follows:

表示句子W的困惑度，N表示句子W中包含的token的数量，

为句子W中包含的N个token，

表示句子W的句子概率。 in,

represents the perplexity of sentence W, N represents the number of tokens contained in sentence W,

is the N tokens contained in the sentence W,

represents the sentence probability of sentence W.

Optionally, for the obtained target dialogue information, correction processing such as adding connectives may be performed, so that the rewritten target dialogue information is of higher quality and smoother.

For example, suppose the target dialogue information is "what is the average height of poker players over 200cm tall", where "height over 200cm" and "poker player" are two key texts, and the connecting word "de" is added between the two key texts , to get the revised "What is the average height of a poker player over 200cm tall".

When applied to a question-and-answer/or dialogue scenario based on table data, after obtaining the target dialogue information rewritten by the dialogue information of the current round, the reply information corresponding to the target dialogue information is obtained through steps S406-S408, and the reply information is fed back.

Step S406: Input the target dialogue information into the language conversion model for conversion processing, and convert the target dialogue information into corresponding SQL statements.

When applied to a question-and-answer/or dialogue scene based on table data, the target dialogue information is input into the language conversion model for conversion processing, and the SQL statement corresponding to the target dialogue information is obtained.

Referring to FIG. 6, FIG. 6 is a framework diagram of Text-to-SQL language conversion based on a context rewriting model provided by an exemplary embodiment. As shown in FIG. 6, based on the multi-round question and answer content shown in FIG. The dialogue information and the two rounds of historical dialogue information of rounds one and two are input into the context rewriting model, and the context rewriting model is used to determine the problem after the rewriting of the dialogue information in this round (that is, the target dialogue information): "The poker player whose height exceeds 200cm What is the average height", input the question and the schema information of the tabular data into the text-to-SQL model, and convert the question into the corresponding SQL statement.

Step S407 , running the SQL statement to obtain the query result by querying the data table storing the table data, determining the reply information according to the query result, and feeding back the reply information.

The server obtains the query result from the data table in which the table data is stored by running the SQL statement corresponding to the target dialogue information.

Further, the server may send the query result to the terminal as the answer information of the question, so that the terminal outputs the answer information for the user to view.

Optionally, based on the query result, the server may also generate reply information according to a preset reply phrase, and feed back the generated reply information to the terminal, so that the terminal outputs the reply information for the user to view.

Optionally, the server may also feed back the query result to the terminal, so that the terminal's preset reply phrases generate reply information, and output the reply information for the user to view.

In this embodiment, the current round of dialogue information and historical dialogue information are input into the context rewriting model, and the context rewriting model is used for classification processing to obtain the rewriting position category and rewriting operation type corresponding to each position in the current round of dialogue information. A candidate position category corresponding to a position; according to each rewriting start position and the corresponding rewriting termination position, determine an object to be rewritten; according to the rewriting operation type of the location of each object to be rewritten, determine the rewriting operation type of the object to be rewritten; according to Each candidate start position and the corresponding candidate end position determine a key text; according to the object to be rewritten and the type of rewriting operation corresponding to the object to be rewritten in the current round of dialogue information, use the key text to rewrite the current round of dialogue information to generate candidate dialogues It can supplement the key information omitted in the current round of dialogue, and replace the reference information with the complete content information referred to, so as to rewrite the current round of dialogue information into high-quality expressions and solve the problems specified in multiple rounds of question and answer. When the omitted questions are applied to the question-and-answer/dialogue scenario based on table data, the corresponding reply information is obtained based on the rewritten target dialogue information, which improves the quality and accuracy of the reply information.

FIG. 7 is a schematic structural diagram of an apparatus for multiple rounds of question and answer provided by an exemplary embodiment of the present application. The apparatus for multiple rounds of question and answer provided by the embodiment of the present application may execute the processing flow provided by the embodiment of the method for multiple rounds of question and answer. As shown in FIG. 7 , the apparatus 70 for multiple rounds of question and answer includes: an information acquisition module 701 , a rewrite prediction module 702 , a rewrite module 703 , a selection module 704 and a reply module 705 .

Specifically, the information acquisition module 701 is configured to, in response to the dialogue request, acquire current dialogue information and historical dialogue information of at least one historical dialogue if it is determined that there is at least one historical dialogue before the current dialogue.

The rewriting prediction module 702 is used to input the current round of dialogue information and historical dialogue information into the context rewriting model, and use the context rewriting model to determine the object to be rewritten in the current round of dialogue information, the rewriting operation type corresponding to the object to be rewritten, and the historical dialogue information. Key text for rewriting.

The rewriting module 703 is configured to rewrite the current round of dialogue information by using key texts according to the to-be-rewritten object and the rewriting operation type corresponding to the to-be-rewritten object in the current round of dialogue information to generate candidate dialogue information.

The selection module 704 is configured to use one of the candidate dialogue information as the target dialogue information after rewriting the current round of dialogue information according to the quality information of the candidate dialogue information.

The reply module 705 is configured to obtain reply information corresponding to the target dialogue information, and feed back the reply information.

Optionally, the rewriting prediction module is applied to input the current round of dialogue information and historical dialogue information into the context rewriting model, and uses the context rewriting model to determine the object to be rewritten in the current round of dialogue information, the type of rewriting operation corresponding to the object to be rewritten, and the history. When the key text used for rewriting in the dialogue information, it is specifically used to: input the current round of dialogue information and historical dialogue information into the context rewriting model, and perform classification processing through the context rewriting model to obtain a classification result containing the following information: In the current round of dialogue information The rewriting position category and the rewriting operation type corresponding to each position, the candidate position category corresponding to each position in the historical dialogue information, wherein the rewriting position category is any of the rewriting start position, the rewriting end position, and the non-rewriting start and end positions; The candidate position category is any one of a candidate start position, a candidate end position, and a non-candidate start and end position. According to the classification result, the object to be rewritten in the current round of dialogue information, the rewriting operation type corresponding to the object to be rewritten, and the key text in the historical dialogue information are determined.

Optionally, when the rewriting prediction module is applied to determine the object to be rewritten in the current round of dialogue information, the type of rewriting operation corresponding to the object to be rewritten, and the key text in the historical dialogue information according to the classification result, it is specifically used for:

According to each rewriting start position and the corresponding rewriting termination position in the classification result, an object to be rewritten is determined; according to the rewriting operation type of the position of each object to be rewritten, the rewriting operation type of the object to be rewritten is determined; The candidate start position and the corresponding candidate end position determine a key text.

Optionally, the object to be rewritten is the position to be inserted or the text to be rewritten, the rewrite operation type includes: insert, delete, replace and keep the same, and the rewrite operation type corresponding to the to-be-inserted position is insert. When the rewriting module is applied to the rewriting operation type corresponding to the object to be rewritten and the object to be rewritten in the dialogue information of the current round, the key text is used to rewrite the dialogue information of the current round, and the candidate dialogue information is generated, and it is specifically used for:

If there is a position to be inserted in the current round of dialogue information, insert at least one key text into the to-be-inserted position; if there is a first text to be rewritten in the current round of dialogue information, and the rewriting operation type of the first text to be rewritten is replace, use at least one text to be rewritten. A key text replaces the first text to be rewritten; if there is a second text to be rewritten in the current round of dialogue information, and the rewriting operation type of the second text to be rewritten is delete, the second text to be rewritten in the current round of dialogue information is deleted.

Optionally, the selection module is specifically used to:

Obtain the sentence probability of each candidate dialogue information, and use the candidate dialogue information with the highest sentence probability as the target dialogue information; or obtain the confusion degree of each candidate dialogue information, and use the candidate dialogue information with the lowest confusion degree as the target dialogue information.

Optionally, the reply module is specifically used to:

Input the target dialogue information into the language conversion model for conversion processing, and convert the target dialogue information into corresponding SQL statements; query the data table storing the table data according to the SQL statements to obtain query results; and determine the reply information according to the query results.

The apparatus provided in this embodiment of the present application may be specifically configured to execute the solution provided by any of the foregoing method embodiments, and specific functions and technical effects that can be achieved will not be repeated here.

FIG. 8 is a schematic structural diagram of an electronic device provided by an exemplary embodiment of the present application. As shown in FIG. 8 , the electronic device 80 includes: a processor 801 , and a memory 802 connected in communication with the processor 801 , and the memory 802 stores computer-executed instructions.

The processor executes the computer-executed instructions stored in the memory to implement the solution provided by any of the above method embodiments, and the specific functions and technical effects that can be achieved are not repeated here. The electronic device may be the server mentioned above.

The embodiments of the present application further provide a computer-readable storage medium, where computer-executable instructions are stored in the computer-readable storage medium, and when the computer-executable instructions are executed by a processor, are used to implement the solution provided by any of the above method embodiments, and the specific functions And the technical effects that can be achieved are not repeated here.

The embodiments of the present application also provide a computer program product, the program product includes: a computer program, the computer program is stored in a readable storage medium, at least one processor of the electronic device can read the computer program from the readable storage medium, at least one The processor executes the computer program to cause the electronic device to execute the solution provided by any of the above method embodiments, and the specific functions and technical effects that can be achieved are not repeated here.

Other embodiments of the present application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses or adaptations of this application that follow the general principles of this application and include common knowledge or conventional techniques in the technical field not disclosed in this application . The specification and examples are to be regarded as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

It is to be understood that the present application is not limited to the precise structures described above and illustrated in the accompanying drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims (8)

1. a method for multiple rounds of question and answer, is characterized in that, comprises: In response to the dialogue request, if it is determined that there is at least one round of historical dialogue before the current round of dialogue, acquiring the current round of dialogue information and the historical dialogue information of the at least one round of historical dialogue; Input the current round of dialogue information and the historical dialogue information into the context rewriting model, and perform classification processing through the context rewriting model to obtain a classification result containing the following information: the rewriting position corresponding to each position in the current round of dialogue information Category and rewriting operation type, the candidate position category corresponding to each position in the historical dialogue information, wherein the rewriting position category is any one of the rewriting start position, the rewriting end position, and the non-rewriting start and end position; the candidate The position category is any of a candidate start position, a candidate end position, and a non-candidate start and end position; According to each rewriting start position and the corresponding rewriting termination position in the classification result, an object to be rewritten is determined, and the object to be rewritten included in the current round of dialogue information is obtained, and the current round of dialogue information includes one or more The object to be rewritten, and the rewrite operation type of the object to be rewritten is determined according to the rewrite operation type of the location of each object to be rewritten, wherein each object to be rewritten is the position to be inserted or the text to be rewritten The type of rewriting operation corresponding to the insertion position is insert, and the type of rewriting operation corresponding to the text to be rewritten can be any of delete, replace, and keep the same; Determine a key text according to each of the candidate start positions and the corresponding candidate end positions in the classification result, and obtain one or more key texts for rewriting in the historical dialogue information; According to the object to be rewritten in the current round of dialogue information and the rewriting operation type corresponding to each object to be rewritten, use the key text to rewrite the current round of dialogue information, and by enumerating possible rewriting methods, Generate possible candidate dialogue information; According to the quality information of the candidate dialogue information, one of the candidate dialogue information is used as the target dialogue information after the current round of dialogue information is rewritten; Acquire reply information corresponding to the target dialogue information, and feed back the reply information. 2. The method according to claim 1, wherein, according to the object to be rewritten in the current round of dialogue information and the rewriting operation type corresponding to the object to be rewritten, using the key text, The round dialogue information is rewritten to generate candidate dialogue information, including: If there is the to-be-inserted position in the current round of dialogue information, insert at least one of the key texts into the to-be-inserted position; If there is a first text to be rewritten in the current round of dialogue information, and the rewriting operation type of the first text to be rewritten is replacement, use at least one of the key texts to replace the first text to be rewritten; If there is a second text to be rewritten in the dialogue information of the current round, and the rewriting operation type of the second text to be rewritten is delete, the second text to be rewritten in the dialogue information of the current round is deleted. 3. The method according to claim 1, wherein, according to the quality information of the candidate dialogue information, using one of the candidate dialogue information as the target dialogue information after the current round of dialogue information is rewritten, comprising: Obtain the sentence probability of each candidate dialogue information, and use the candidate dialogue information with the highest sentence probability as the target dialogue information; or, The perplexity degree of each candidate dialogue information is obtained, and the candidate dialogue information with the lowest perplexity degree is used as the target dialogue information. 4. The method according to claim 1, wherein the acquiring the current round of dialogue information and the historical dialogue information of the at least one round of historical dialogues comprises: Obtain the dialogue text of the current round and the historical dialogue text of the at least one round of historical dialogue; The current round of dialogue texts and the historical dialogue texts of each round of historical dialogues are processed by word segmentation, respectively, to obtain the token sequence of the current round of dialogue texts and the token sequence of the historical dialogue texts of each round of historical dialogues. 5. The method according to claim 1, wherein the acquiring the reply information corresponding to the target dialogue information comprises: Inputting the target dialogue information into a language conversion model for conversion processing, and converting the target dialogue information into a corresponding SQL statement; Query the data table storing the tabular data according to the SQL statement to obtain the query result; Reply information is determined according to the query result. 6. A device for multi-round question and answer, characterized in that, comprising: an information acquisition module, configured to, in response to the dialogue request, acquire the dialogue information of the current round and the historical dialogue information of the at least one round of historical dialogue if it is determined that there is at least one round of historical dialogue before the current round of dialogue; The rewriting prediction module is used to input the current round of dialogue information and the historical dialogue information into the context rewriting model, and perform classification processing through the context rewriting model to obtain a classification result containing the following information: A rewrite location type and rewrite operation type corresponding to a location, a candidate location class corresponding to each location in the historical dialogue information, wherein the rewrite location class is any of the rewrite start location, rewrite end location, and non-rewrite start and end locations. One; the candidate position category is any one of a candidate start position, a candidate end position, and a non-candidate start and end position; according to each of the rewrite start positions and the corresponding rewrite termination positions in the classification results, determine a waiting Rewriting objects, obtaining the objects to be rewritten included in the current round of dialogue information, where the current round of dialogue information includes one or more objects to be rewritten, and determining the rewriting operation type according to the location of each object to be rewritten The rewriting operation type of the object to be rewritten, wherein each object to be rewritten is the position to be inserted or the text to be rewritten, the rewriting operation type corresponding to the position to be inserted is insert, and the rewriting operation type corresponding to the text to be rewritten can be delete, replace , remain as it is; determine a key text according to each of the candidate starting positions and the corresponding candidate ending positions in the classification result, and obtain one or more key texts for rewriting in the historical dialogue information ; The rewriting module is used to rewrite the current round of dialogue information by using the key text according to the object to be rewritten in the current round of dialogue information and the corresponding rewriting operation type of each object to be rewritten, and enumeration Possible rewriting methods to generate possible candidate dialogue information; a selection module, configured to use one of the candidate dialogue information as the target dialogue information after rewriting the current round of dialogue information according to the quality information of the candidate dialogue information; A reply module, configured to obtain reply information corresponding to the target dialogue information, and feed back the reply information. 7. An electronic device, comprising: a processor, and a memory communicatively connected to the processor; the memory stores computer-executable instructions; The processor executes computer-implemented instructions stored in the memory to implement the method of any of claims 1-5. 8. A computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, is used to implement any one of claims 1-5 method described in item.

Images