CN114637765A

CN114637765A - Man-machine interaction method, device and equipment based on form data

Info

Publication number: CN114637765A
Application number: CN202210440950.0A
Authority: CN
Inventors: 惠彬原; 黎槟华; 李永彬; 孙健
Original assignee: Alibaba Damo Institute Hangzhou Technology Co Ltd
Current assignee: Alibaba Damo Institute Hangzhou Technology Co Ltd
Priority date: 2022-04-26
Filing date: 2022-04-26
Publication date: 2022-06-17

Abstract

The application provides a man-machine interaction method, a man-machine interaction device and man-machine interaction equipment based on table data. According to the method, a search base is constructed in an off-line mode, and SQL sentences and problem representation vectors corresponding to the SQL sentences, which are supported and inquired by table data, are stored in the search base. When the questions are asked and answered/dialogues online, the input questions are coded to generate semantic representation vectors of the input questions, similarity matching is conducted on the semantic representation vectors of the input questions and question representation vectors in a search library, and then SQL sentences which are accurately matched with the input questions can be found, Text-to-SQL language conversion is achieved, model training is not needed to be conducted on the basis of a data set in a specific application field, generalization is stronger, and efficiency is higher; furthermore, SQL sentences corresponding to the input questions are operated, namely, the table data can be inquired to obtain answer information of the input questions, the response time during on-line question answering/conversation can be shortened, and the response timeliness of the man-machine interaction system based on the table data is improved.

Description

Man-machine interaction method, device and equipment based on form data

Technical Field

The application relates to the field of artificial intelligence, in particular to a man-machine interaction method, a man-machine interaction device and man-machine interaction equipment based on form data.

Background

A Table data based human-computer interaction system (Table QA) is a function of converting a natural language problem input by a user into an SQL statement, finding a result corresponding to the SQL statement in a data Table for storing Table data in a database and feeding back the result to the user. Because the form is a common data storage structure in daily work and life, the man-machine interaction system based on the form data has wide application scenes, such as data query, statistics, screening and the like, and can be applied to various application fields of government affairs, finance, energy and the like.

At present, a man-machine interaction system based on table data converts a natural language problem input by a user into a corresponding SQL statement by using a language conversion model, queries a data table in a database to obtain a query result corresponding to the SQL statement, and feeds the query result back to the user. However, such a language conversion model that directly converts a natural language problem into a corresponding SQL statement has poor domain-crossing generalization, and when applied to different application domains, a training set in a specific domain needs to be constructed for model training.

Disclosure of Invention

The application provides a form data-based human-computer interaction method, device and equipment, which are used for solving the problem that a language conversion model used by a current form data-based human-computer interaction system directly converts a natural language problem into a corresponding SQL statement is poor in cross-domain generalization.

In one aspect, the application provides a human-computer interaction method based on table data, including:

responding to a question-answering request based on table data, encoding an input question, and generating a semantic representation vector of the input question;

searching a problem representation vector matched with the semantic representation vector in a search library according to the semantic representation vector, wherein SQL sentences supporting query of the table data and the problem representation vector corresponding to each SQL sentence are stored in the search library, and the problem representation vector corresponding to the SQL sentences refers to the semantic representation vector of the natural language problem corresponding to the SQL sentences;

operating SQL sentences corresponding to the question representation vectors matched with the semantic representation vectors to query the table data to obtain answer information of the input question;

and feeding back the answer information.

In another aspect, the present application provides a human-computer interaction device based on form data, including:

the system comprises an encoding module, a semantic representation module and a semantic representation module, wherein the encoding module is used for responding to a question-answer request based on table data, encoding an input question and generating a semantic representation vector of the input question;

the retrieval module is used for searching a problem representation vector matched with the semantic representation vector in a retrieval library according to the semantic representation vector, wherein the retrieval library stores SQL sentences supporting query of the table data and the problem representation vector corresponding to each SQL sentence, and the problem representation vector corresponding to the SQL sentence refers to the semantic representation vector of the natural language problem corresponding to the SQL sentence;

the query module is used for operating SQL sentences corresponding to the question representation vectors matched with the semantic representation vectors so as to query the table data to obtain answer information of the input question;

and the feedback module is used for feeding back the answer information.

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

the memory stores computer-executable instructions;

the processor executes computer execution instructions stored in the memory to realize the man-machine interaction method based on the table data.

In another aspect, the present application provides a computer-readable storage medium, in which computer-executable instructions are stored, and the computer-executable instructions are executed by a processor to implement the above-mentioned human-computer interaction method based on table data.

According to the man-machine interaction method, the man-machine interaction device and the man-machine interaction equipment based on the tabular data, a search library is constructed in advance in an off-line mode, SQL sentences supporting query of the tabular data and problem representation vectors corresponding to the SQL sentences are stored in the search library, and high-quality SQL sentences and corresponding problem representation vectors are provided. During on-line question answering/dialogue, the semantic representation vector of the input question is generated by encoding the input question, similarity matching is carried out on the semantic representation vector of the input question and the question representation vector in the search library based on the search library constructed off-line, so that an SQL sentence accurately matched with the input question can be found in the search library, Text-to-SQL language conversion is realized, model training based on a specific application field is not needed, and the generalization of the Text-to-SQL language conversion method is stronger; the Text-to-SQL language conversion method based on the retrieval formula has higher efficiency; furthermore, the SQL sentence corresponding to the input question is operated, the table data can be inquired to obtain answer information of the input question, response time during on-line question answering/conversation can be shortened, and response timeliness of a man-machine interaction system based on the table data is improved.

Drawings

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

FIG. 1 is a schematic diagram of a network architecture on which the present application is based;

FIG. 2 is a flowchart of a method for human-computer interaction based on tabular data according to an exemplary embodiment of the present application;

FIG. 3 is a flowchart of a method for constructing a search corpus, according to an exemplary embodiment of the present application;

FIG. 4 is an exemplary diagram of a data table provided by an exemplary embodiment of the present application;

FIG. 5 is a diagram illustrating a tree structure of an example SQL statement provided by an exemplary embodiment of the present application;

FIG. 6 is a general flowchart of a Text-to-SQL conversion according to an embodiment of the present application;

FIG. 7 is a schematic structural diagram of a human-computer interaction device based on tabular data according to an exemplary embodiment of the present application;

FIG. 8 is a schematic structural diagram of a human-computer interaction device based on tabular data according to another exemplary embodiment of the present application;

fig. 9 is a schematic structural diagram of an electronic device according to an example embodiment of the present application.

With the above figures, there are shown specific embodiments of the present application, which will be described in more detail below. These drawings and written description are not intended to limit the scope of the inventive concepts in any manner, but rather to illustrate the inventive concepts to those skilled in the art by reference to specific embodiments.

Detailed Description

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

The terms referred to in this application are explained first:

data table: the data storage method includes storing data in a data table in a Structured Query Language (SQL) database, completing permanence of the data to be stored after the data is stored in a corresponding data table of the database, and acquiring data content of the data by accessing the data table of the Query database.

Data elements: also referred to as a schema of the data table, refers to information such as table names, column names, values, etc. in the data table.

Table QA: a man-machine interaction system based on table data (or table knowledge) achieves the functions of converting natural language questions input by users into SQL sentences, finding results corresponding to the SQL sentences in a data table used for storing the table data in a database and feeding back the results to the users. Such as a question-answering/dialogue system based on tabular data.

Text-to-SQL language conversion model: is a core technology of a Table data-based human-computer interaction system (Table QA), a language understanding (semantic parsing) way, which is used for converting a human natural language described question (Text) into a computer executable Structured Query Language (SQL). The model can realize free interaction between a person and a form/database without the need of learning complex SQL grammar by a user. Further, a Table/database based question answering/dialogue system (Table QA) may be implemented in dependence on the model.

BERT: all called Bidirectional Encoder reproduction from transformations, is a pre-training model. This model has shown surprising performance in a machine reading comprehension level test: both metrics surpass humans in all and create sota (state of the art) performance in 11 different Natural Language Processing (NLP) tests. The input of the model is natural language sentences, and the output is semantic representation vectors of the natural language sentences.

RoBERTA: the method is a strictly optimized BERT pre-training model which is built on the basis of BERT and modifies partial hyper-parameters.

Furthermore, the terms "first", "second", etc. are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. In the description of the following examples, "plurality" means two or more unless specifically limited otherwise.

Aiming at the problems that in the existing man-machine interaction system based on table data, a language conversion model which directly converts a natural language problem into a corresponding SQL statement is poor in cross-domain generalization and needs to be retrained if a specific vertical domain scene is changed, the method is based on a new retrieval type Text-to-SQL language conversion method to realize man-machine interaction based on the table data.

Specifically, a search base is constructed based on offline data: based on mode information of a data table where table data used by a current interactive system is located, possible SQL sentences used for inquiring the data table are generated through sampling, the SQL sentences are converted into corresponding natural language problems, semantic representation vectors of the natural language problems are generated through coding the natural language problems and serve as problem feature vectors corresponding to the SQL sentences, the problem feature vectors corresponding to the SQL sentences are stored in a database in a key-value pair (key-value) mode, and construction of a search base is completed. When man-machine interaction is carried out on line, the user input problem is coded to generate a semantic representation vector of the input problem, based on the similarity of the representation vectors, the problem representation vector matched with the semantic representation vector of the input problem is searched in a search library, and an SQL sentence corresponding to the problem representation vector is used as the SQL sentence corresponding to the input problem. And querying the data table based on the SQL sentence corresponding to the input question, wherein the query result is used as answer information of the input question.

When the method is applied to different application fields and table data are changed, a search base is reconstructed based on the new table data, data sets in different application fields do not need to be constructed for model training, the generalization of the Text-to-SQL language conversion method is improved, and the problem of poor generalization is solved.

In addition, when human-computer interaction is carried out on line, based on the semantic representation vector of the input problem, the SQL sentence corresponding to the input problem is inquired in the search library, and compared with the prior art that a Text-to-SQL language conversion model is used for converting the natural language problem input by the user into the corresponding SQL sentence, the time consumption is shortened, and the system response time can be shortened.

The man-machine interaction method based on the form data can be applied to a man-machine interaction system based on the form data (form knowledge or a database) in various application fields, such as a question-answering/dialogue system based on the form data.

Fig. 1 is a schematic diagram of a network architecture based on the present application, and the network architecture shown in fig. 1 may specifically include a server and a terminal.

The server may be a server cluster arranged at the cloud, and the server constructs a search library based on the offline data. The problem representation vector and the corresponding SQL statement are stored in the search base. The server stores a data table of a database, and the data table stores table data which needs to be inquired when question answering/conversation is carried out.

The terminal may specifically be an electronic device having a network communication function, an operation function, and an information display function, and includes, but is not limited to, a smart phone, a tablet computer, a desktop computer, an internet of things device, and the like.

The terminal submits a question-answer request containing an input question to the server based on user operation. The server encodes the input problem to generate a semantic representation vector of the input problem; and querying a retrieval library based on the semantic representation vector of the input question, determining an SQL (structured query language) statement corresponding to the input question, querying a data table based on the SQL statement to obtain answer information of the input question, and feeding back the answer information to the terminal. And the terminal receives answer information fed back by the server. The terminal can output answer information for the user to view.

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

Fig. 2 is a flowchart of a man-machine interaction method based on table data according to an exemplary embodiment of the present application. The man-machine interaction method based on the table data provided by the embodiment can be specifically applied to the aforementioned server. As shown in fig. 2, the method comprises the following specific steps:

step S201, responding to a question-answer request based on table data, encoding an input question and generating a semantic representation vector of the input question.

The server responds to a question-answer request of the terminal, wherein the question-answer request comprises questions input by a user through the terminal, extracts the input questions, encodes the input questions and generates semantic representation vectors of the input questions.

In this step, the input question may be input into a pre-trained coding model for coding, and a semantic representation vector of the input question is generated through coding of the coding model.

Optionally, the coding model for coding the input problem to generate the characterization vector may specifically be implemented by models such as bert (bidirectional Encoder reconstruction from transforms), RoBERTa, and the like, and may also be implemented by other models having similar functions, which is not specifically limited in this embodiment.

For example, the input question may be input into a pre-trained BERT model, and a semantic representation vector of the input question may be generated by BERT model encoding.

Step S202, according to the semantic representation vector of the input problem, searching a problem representation vector matched with the semantic representation vector of the input problem in a search library, and determining the SQL sentence corresponding to the problem representation vector as the SQL sentence corresponding to the input problem, wherein the SQL sentence supporting query by table data and the problem representation vector corresponding to each SQL sentence are stored in the search library, and the problem representation vector corresponding to the SQL sentence refers to the semantic representation vector of the natural language problem corresponding to the SQL sentence.

In this embodiment, the search library is previously constructed on the server. The search library stores a plurality of pieces of data, each piece of data comprises an SQL statement and a problem representation vector corresponding to the SQL statement. And querying a data table for storing table data when each SQL statement runs to obtain a query result. The search base comprises a plurality of SQL sentences of which the data tables support query, and the possible SQL sentences are covered as much as possible. Therefore, for any input problem of a user, the corresponding SQL statement can be accurately found in the search library.

The problem representation vector corresponding to the SQL statement is a semantic representation vector generated by coding a natural language problem corresponding to the SQL statement.

Optionally, in this step, the similarity between the semantic feature vector of the input question and each question feature vector in the search base may be calculated, and the question feature vector having the highest similarity with the semantic feature vector of the input question is determined as the question feature vector matching the semantic feature vector of the input question.

Alternatively, the similarity of the two semantic representation vectors may be determined by calculating the cosine similarity of the two semantic representation vectors. For example, the similarity of two semantic representation vectors can be determined by the following formula one:

formula one

A and B respectively represent two semantic representation vectors, and similarity represents cosine similarity of A and B.

Cosine similarity measures the similarity between two vectors by measuring their cosine values of their angle. The cosine value of the 0-degree angle is 1, and the cosine value of any other angle is not more than 1; and its minimum value is-1. The cosine of the angle between the two vectors thus determines whether the two vectors point in approximately the same direction. When the two vectors have the same direction, the cosine similarity value is 1; when the included angle of the two vectors is 90 degrees, the value of the cosine similarity is 0; the cosine similarity has a value of-1 when the two vectors point in completely opposite directions. The result is independent of the length of the vector, only the pointing direction of the vector. Cosine similarity is commonly used in the forward space, so the similarity of two semantic representation vectors has a value between 0 and 1.

Optionally, in this step, a distance between the semantic feature vector of the input question and each question feature vector in the search library may be calculated, and the question feature vector having the smallest distance from the semantic feature vector of the input question is determined as the question feature vector matching the semantic feature vector of the input question. Wherein, the distance between the two semantic representation vectors can be euclidean distance (or euclidean distance), etc.

In addition, the natural language question corresponding to the SQL sentence and the input question of the user are coded by adopting the same coding mode, so that the question representation vector matched with the semantic representation vector of the input question can be accurately searched in a search library based on the similarity/difference of the representation vectors, the matching accuracy is improved, and the question-answering/dialogue quality based on the table data is improved.

After the problem representation vector matched with the semantic representation vector is found in the search library, the SQL sentence corresponding to the problem representation vector matched with the semantic representation vector is the SQL sentence corresponding to the input problem, the SQL sentence corresponding to the problem representation vector matched with the semantic representation vector is determined as the SQL sentence corresponding to the input problem, and Text-to-SQL language conversion is achieved.

Step S203, running the SQL statement corresponding to the input question to query the table data to obtain answer information of the input question.

After SQL sentences corresponding to the input problems are determined and Text-to-SQL language conversion is achieved, the SQL sentences corresponding to the input problems are operated in the database, query results corresponding to the SQL sentences are obtained by querying the data table used for storing table data, and answer information of the input problems is obtained.

And step S204, feeding back answer information.

After the answer information of the input question is obtained through inquiry, the server feeds the answer information of the input question back to the terminal equipment.

In the embodiment, a search library is constructed in advance in an offline manner, and SQL statements and problem representation vectors corresponding to each SQL statement, which are supported by table data, are stored in the search library, so that high-quality SQL statements and corresponding problem representation vectors are provided. During on-line question answering/dialogue, the input question is coded to generate a semantic representation vector of the input question, similarity matching is carried out on the semantic representation vector of the input question and a question representation vector in a search library based on the search library constructed off-line, an SQL sentence accurately matched with the input question can be found in the search library, Text-to-SQL language conversion is realized, model training based on a specific application field is not needed, and the generalization of the Text-to-SQL language conversion method is stronger; the Text-to-SQL language conversion method based on the retrieval formula has higher efficiency; furthermore, SQL sentences corresponding to the input questions are operated, namely, the table data can be inquired to obtain answer information of the input questions, the response time during on-line question answering/dialogue can be shortened, and the response timeliness of the man-machine interaction system based on the table data is improved.

In an alternative embodiment, the server may build a search base based on offline data in advance before conducting online question answering/dialogue. The following describes in detail a specific process of the server building a search library based on offline data in advance.

Specifically, the server obtains SQL statements supporting query of table data and a problem representation vector corresponding to each SQL statement; and correspondingly storing the problem representation vector corresponding to each SQL statement and the SQL statement to construct a search library.

The constructed search library stores a plurality of pieces of data, each piece of data comprises an SQL statement and a problem representation vector corresponding to the SQL statement. And querying a data table for storing table data when each SQL statement runs to obtain a query result. The search base comprises a plurality of SQL sentences of which the data tables support query, and the possible SQL sentences are covered as much as possible. Therefore, for any input problem of a user, the corresponding SQL statement can be accurately found in the search library.

Referring to fig. 3, fig. 3 is a flowchart of a method for constructing a search base according to an exemplary embodiment of the present application, and as shown in fig. 3, the method for constructing a search base includes the following specific steps:

step S301, acquiring mode information of a data table for storing table data; and generating an SQL statement supporting query of the data table according to the mode information of the data table.

In a practical application scenario, the table data based on the question answering/dialogue system may include one or more tables, and the content data of the table is stored in a data table in the database.

In this step, the server may extract the data element from the header, column name, and value of the data table as the schema information of the data table. Further, based on the preconfigured SQL statement generation rule, a plurality of SQL statements for querying the data table are generated.

Specifically, the SQL statement generation rule may include a plurality of statement templates of the SQL statement, and the statement templates of the SQL statement may include positions to be filled, where the positions to be filled are used to fill data elements in the data table. And filling the data elements in the positions to be filled in the sentence pattern template of the SQL sentence to form the SQL sentence. And filling different data elements in the positions to be filled in the sentence pattern template of the SQL sentences to form a plurality of different SQL sentences.

The SQL syntax rules of an SQL database are as follows:

SQLs ::= SQL | SQL intersect SQLs | SQL union SQLs | …

SQL ::= Select | Select Where | Select Order | Select Order Fliter | …

Select ::= SELECT A | SELECT AA |…

Where ::= Where Conditions

Conditions ::= A op value | A op SQL

C ::= table | table.column

op ::= == | != | > | >= | < | <= | like | in | nor in

where "|" means "or" means ":: = means assigned value, i.e. the object on the left side can have several different expressions on the right side. Table refers to a table name of a data table, column refers to a column name in the data table, and value may refer to a value in the data table.

Illustratively, one sentence pattern template may be: SQL = SELECT | SELECTWHERE | SELECT order, three of which form a tree structure, i.e. the SELECT part is the parent node of the SELECT WHERE part, and there is SELECT WHERE part only when the SELECT part exists; SELECT WHERE is the parent node of the SELECT ORDER part, which will only exist if part SELECT WHERE exists. SELECT represents a query action; SELECT WHERE denotes the object of the query; the SELECT ORDER represents the ordering of the queries.

Specifically, in section SELECT WHERE WHERE = WHERE Conditions, WHERE Conditions are used to represent query objects. The Conditions of the Where Conditions section may be: the value of A op, A is used to represent the calculation function required by the query, and has the form of "C", "MIN C", "MAX C", "AVG C", "COUNT C", "SUM C", etc., where C is a data element of the data table and may be a table name or a column name of the data table. MIN, MAX, AVG, COUNT are different calculation functions. op is a comparison operator and has the form "=", "| =", "> =", "<=", "like", "in", "nor in", and the like. The value is a value used to indicate comparison, is a data element of the data table, and may be a value in the data table.

For example, taking the data table shown in fig. 4 as an example, the name of the data table is an item information table, and the data table includes the following: project name, performer, project level, and city of the project. The data table supports the following SQL statements: the city WHERE the SELECT COUNT project name FROM project information table WHERE = = = "beijing", and the tree structure of this SQL statement is as shown in fig. 5.

In this embodiment, the data elements include three different types, namely, a table name, a column name, and a value, and when a position to be filled in a schema template of an SQL statement is filled with a data element, the data element of the corresponding type needs to be filled based on the type of the data element that the position to be filled allows to be filled.

Optionally, for each sentence pattern template, multiple SQL sentences may be obtained by randomly filling multiple data elements to respective positions to be filled in the sentence pattern template according to a random condition.

For example, assume that the table name in the data element extracted from the data table is: a student. The column names in the extracted data elements include: name, height, weight, age. Some of the extracted data elements have values: years: 10. 12; height: 160. 155. One sentence pattern template is:

SELECT（table. column） FROM（table） WHERE （table. column）> （value）。

wherein "()" denotes the position to be filled. Then, randomly filling the data elements to the corresponding positions to be filled in the sentence pattern template, and obtaining at least the following SQL sentences:

SELECT MAX (height) FROM (student) WHERE (years) > (10)// characterizing the height of students with query years greater than 10 years of age and highest height;

SELECT (name) FROM (student) WHERE (height) > (155)// name characterizing the student whose query height is greater than 175.

Optionally, in this step, according to the mode information of the data table, a context-free grammar may be used to recursively enumerate the data elements to be filled at each position to be filled in the sentence template, and as many SQL sentences as possible are generated in an enumeration manner.

Step S302, each SQL statement is converted into a corresponding natural language question.

After the SQL statements supporting the query by the data table are obtained, each SQL statement is converted into a corresponding natural language problem.

Specifically, the SQL sentences can be input into a pre-trained SQL-to-Text conversion model for coding, and the SQL sentences are converted into corresponding natural language problems through the SQL-to-Text conversion model, so that the high-quality natural language problems corresponding to the SQL sentences can be obtained automatically in batch.

Alternatively, the SQL-to-Text conversion model may adopt BART (Bidirectional and Auto-regenerative transformations), T5 (Transfer Text-to-Text transformation), and other models with similar functions, and this embodiment is not limited in detail here.

In this embodiment, the SQL-to-Text conversion model used to convert each SQL statement into a corresponding natural language question is a pre-trained model, and can be obtained by pre-training on a public data set.

Step S303, the natural language problem corresponding to each SQL statement is coded, and a problem representation vector corresponding to each SQL statement is generated.

After each SQL statement is converted into a corresponding natural language problem, the natural language problem corresponding to each SQL statement is coded to generate a corresponding problem representation vector.

Illustratively, the input question may be input into a pre-trained BERT model in this step, and a semantic representation vector of the input question is generated by BERT model coding.

For example, the semantic representation vector generated by the BERT model may be a vector with 768 dimensions, and the dimensions of the semantic representation vector may be adjusted based on the BERT model, which is not specifically limited herein.

In this embodiment, the input question is encoded during the on-line question answering/dialogue in step S201, and the encoding method is the same as that used to encode the natural language question corresponding to each SQL statement in step S303.

Illustratively, the natural language question corresponding to each SQL statement is input into the coding model for coding, and a question representation vector corresponding to each SQL statement is generated. When in on-line question answering/conversation, the input questions are input into the same coding model for coding, and semantic representation vectors of the input questions are generated. Therefore, when on-line question answering/dialogue is carried out, the question representation vectors matched with the semantic representation vectors of the input questions can be accurately searched in the search library based on the similarity/difference of the representation vectors, the matching accuracy is improved, and the question answering/dialogue quality based on the table data is improved.

Through the steps S301 to S303, the SQL statements that the table data supports the query and the problem representation vector corresponding to each SQL statement are obtained.

After the SQL statements supporting the query by the table data and the problem representation vector corresponding to each SQL statement are obtained, the problem representation vector corresponding to each SQL statement and the SQL statement are stored correspondingly in step S304, and a search library is constructed.

Step S304, the problem representation vector corresponding to each SQL statement is used as a key, the SQL statement is used as a value of the key, and the problem representation vector corresponding to each SQL statement and the SQL statement are stored in a database in a key value pair mode to construct a search library.

In this embodiment, the problem representation vector and the SQL statement corresponding to each SQL statement may be stored in a key-value pair (key-value) manner to form a key-value type database, and the construction of the search library is completed. That is, the search library is a key-value type database, and stores a plurality of key value pairs, where the key (key) of each key value pair is a problem characterization vector corresponding to an SQL statement, and the value (value) of the key (key) is the SQL statement.

Illustratively, the problem representation vector corresponding to the SQL statement and the SQL statement are stored in the search base in a format of < problem representation vector, SQL statement >.

Based on the search library, the semantic representation vector of the input problem is matched with the key (key) of the key value pair in the search library, and the value (value) corresponding to the input problem is determined, so that the SQL sentence corresponding to the input problem can be obtained, and the matching efficiency can be improved.

In other embodiments of this embodiment, the problem representation vector corresponding to each SQL statement and the SQL statement may be stored correspondingly in other manners to construct the search library.

For example, the problem characterization vector corresponding to the SQL statement and the SQL statement may be stored in a data table of a relational database, and the data table may include two columns, where one column is the SQL statement and the other column is the problem characterization vector corresponding to the SQL statement.

Optionally, the search library may further store a natural language problem corresponding to the SQL statement, and store mapping relationships between the SQL statement and the natural language problem corresponding to the SQL statement, and a characterization vector of the natural language problem.

Referring to fig. 6, fig. 6 is a general flowchart of Text-to-SQL conversion according to an embodiment of the present application. As shown in fig. 6, in the offline stage, an SQL statement that the data table supports query is generated based on table mode sampling, a natural language problem corresponding to the SQL statement is generated through an SQL-to-Text conversion model, the natural language problem corresponding to the SQL statement is encoded through an encoding model to generate a problem representation vector, and a search library is constructed, in which the SQL statement and the problem representation vector corresponding to the SQL statement are stored. In the on-line question answering/dialogue stage, questions input by a user are coded through a coding model to generate corresponding semantic representation vectors; and searching in a search library based on the semantic representation vector of the input question, and determining SQL corresponding to the input question.

In this embodiment, the server obtains mode information of a data table for storing table data; generating an SQL statement supporting query of the data table according to the mode information of the data table; converting each SQL statement into a corresponding natural language problem, coding the natural language problem corresponding to each SQL statement, generating a problem representation vector corresponding to each SQL statement, storing the problem representation vector corresponding to each SQL statement and the SQL statement into a database in a key value pair mode to construct a search library, thereby constructing the search library in advance based on offline data, wherein the search library comprises a plurality of SQL statements supported and queried by a data table and covers the SQL statements as much as possible; based on an off-line constructed search library, similarity matching is carried out on semantic representation vectors of input problems and problem representation vectors in the search library, SQL sentences which are accurately matched with the input problems can be found in the search library, Text-to-SQL language conversion is realized, model training based on a specific application field is not needed, and the generalization of a Text-to-SQL language conversion method is stronger; and the Text-to-SQL language conversion method based on the retrieval formula has higher efficiency.

Fig. 7 is a schematic structural diagram of a human-computer interaction device based on tabular data according to an exemplary embodiment of the present application. The man-machine interaction device based on the form data, provided by the embodiment of the application, can execute the processing flow provided by the man-machine interaction method based on the form data. As shown in fig. 7, the human-computer interaction device 70 based on table data includes: an encoding module 701, a retrieval module 702, a query module 703 and a feedback module 704.

Specifically, the encoding module 701 is configured to encode the input question in response to a question-answering request based on the table data, and generate a semantic representation vector of the input question.

The retrieval module 702 is configured to search, according to the semantic representation vectors, problem representation vectors matched with the semantic representation vectors in a retrieval library, where the retrieval library stores SQL statements that the table data supports queries and problem representation vectors corresponding to each SQL statement, and the problem representation vectors corresponding to the SQL statements refer to the semantic representation vectors of the natural language problem corresponding to the SQL statements.

The query module 703 is configured to run an SQL statement corresponding to the question representation vector matched with the semantic representation vector to query the table data to obtain answer information of the input question.

And a feedback module 704 for feeding back answer information.

The apparatus provided in the embodiment of the present application may be specifically configured to execute the scheme provided in the embodiment of the method corresponding to fig. 2, and specific functions and technical effects that can be achieved are not described herein again.

In this embodiment, a search library is pre-constructed offline, and SQL statements and problem representation vectors corresponding to each SQL statement, which are supported by table data, are stored in the search library, so that high-quality SQL statements and corresponding problem representation vectors are provided. During on-line question answering/dialogue, the input question is coded to generate a semantic representation vector of the input question, similarity matching is carried out on the semantic representation vector of the input question and a question representation vector in a search library based on the search library constructed off-line, an SQL sentence accurately matched with the input question can be found in the search library, Text-to-SQL language conversion is realized, model training based on a specific application field is not needed, and the generalization of the Text-to-SQL language conversion method is stronger; the Text-to-SQL language conversion method based on the retrieval formula has higher efficiency; furthermore, SQL sentences corresponding to the input questions are operated, namely, the table data can be inquired to obtain answer information of the input questions, the response time during on-line question answering/dialogue can be shortened, and the response timeliness of the man-machine interaction system based on the table data is improved.

Referring to fig. 8, fig. 8 is a schematic structural diagram of a human-computer interaction device based on table data according to another exemplary embodiment of the present application, and based on the device provided in the foregoing embodiment, as shown in fig. 8, the human-computer interaction device 70 based on table data may further include: the search library construction module 705.

The search library construction module 705 comprises:

and the SQL sampling sub-module is used for acquiring SQL sentences which support query by the table data and problem representation vectors corresponding to each SQL sentence.

And the search library construction submodule is used for correspondingly storing the problem representation vector corresponding to each SQL statement and the SQL statement to construct a search library.

Optionally, the SQL sampling sub-module is further configured to:

acquiring mode information of a data table for storing table data, and generating an SQL statement of a data table supporting query according to the mode information of the data table; converting each SQL statement into a corresponding natural language question; and coding the natural language problem corresponding to each SQL statement to generate a problem representation vector corresponding to each SQL statement.

Optionally, the SQL sampling sub-module is further configured to:

acquiring a sentence pattern template of the SQL sentence, wherein the sentence pattern template comprises at least one item of position to be filled; and filling the data elements of the data table with the positions to be filled in the sentence pattern template according to the pattern information of the data table to obtain the SQL sentences of which the table data supports query, wherein the model information comprises a plurality of data elements of the data table.

Optionally, the search library construction sub-module is further configured to:

and taking the problem representation vector corresponding to each SQL statement as a key, taking the SQL statement as a value of the key, and storing the problem representation vector corresponding to each SQL statement and the SQL statement into a database in a key-value pair mode to construct a search library.

Optionally, the SQL sampling sub-module is further configured to:

and inputting the natural language question corresponding to each SQL statement into a coding model for coding, and generating a question representation vector corresponding to each SQL statement.

Further, the encoding module is further configured to: and inputting the input question into the coding model for coding to generate a semantic representation vector of the input question.

Optionally, the retrieval module is further configured to:

and calculating the similarity between the semantic representation vector and each problem representation vector in the search library, and determining the problem representation vector with the highest similarity with the semantic representation vector as the problem representation vector matched with the semantic representation vector.

The apparatus provided in the embodiment of the present application may be specifically configured to execute the scheme provided in any one of the method embodiments, and specific functions and technical effects that can be achieved are not described herein again.

Fig. 9 is a schematic structural diagram of an electronic device according to an example embodiment of the present application. As shown in fig. 9, the electronic apparatus 90 includes: a processor 901, and a memory 902 communicatively coupled to the processor 901, the memory 902 storing computer-executable instructions.

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

The embodiment of the present application further provides 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, the computer-executable instructions are used to implement the solutions provided in any of the above method embodiments, and specific functions and technical effects that can be achieved are not described herein again.

An embodiment of the present application further provides a computer program product, where the program product includes: 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, and the at least one processor executes the computer program to enable the electronic device to execute the scheme provided by any one of the above method embodiments, and specific functions and achievable technical effects are not described herein again.

Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the application being indicated by the following claims.

It will be understood that the present application is not limited to the precise arrangements that have been described above and shown in the 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

1. A man-machine interaction method based on table data is characterized by comprising the following steps:

responding to a question-answer request based on table data, encoding an input question, and generating a semantic representation vector of the input question;

according to the semantic representation vectors, searching problem representation vectors matched with the semantic representation vectors in a search library, and determining SQL sentences corresponding to the problem representation vectors as SQL sentences corresponding to the input problems, wherein the search library stores SQL sentences supporting query by the table data and problem representation vectors corresponding to each SQL sentence, and the problem representation vectors corresponding to the SQL sentences refer to the semantic representation vectors of the natural language problems corresponding to the SQL sentences;

operating SQL sentences corresponding to the input questions to query the table data to obtain answer information of the input questions;

and feeding back the answer information.

2. The method of claim 1, wherein prior to finding a problem token vector in a search corpus that matches the semantic token vector based on the semantic token vector, further comprising:

acquiring SQL sentences which are queried by the table data support and problem representation vectors corresponding to each SQL sentence;

and correspondingly storing the problem representation vector corresponding to each SQL statement and the SQL statement to construct the search library.

3. The method of claim 2, wherein the obtaining the SQL statements and the problem characterization vectors corresponding to each SQL statement of the tabular data support query comprises:

acquiring mode information of a data table for storing the table data, and generating an SQL statement supporting query of the data table according to the mode information of the data table;

converting each SQL statement into a corresponding natural language question;

and coding the natural language problem corresponding to each SQL statement to generate a problem representation vector corresponding to each SQL statement.

4. The method according to claim 3, wherein the generating the SQL statement that the data table supports query according to the schema information of the data table comprises:

acquiring a sentence pattern template of the SQL sentence, wherein the sentence pattern template comprises at least one item of position to be filled;

and filling the positions to be filled in the sentence pattern template into the data elements of the data table according to the pattern information of the data table to obtain the SQL sentences of which the table data supports query, wherein the model information comprises a plurality of data elements of the data table.

5. The method according to claim 2, wherein said storing the problem characterization vector corresponding to each SQL statement in correspondence with the SQL statement, and constructing the search base, comprises:

and taking the problem representation vector corresponding to each SQL statement as a key, taking the SQL statement as a value of the key, and storing the problem representation vector corresponding to each SQL statement and the SQL statement into a database in a key-value pair mode to construct the search library.

6. The method according to claim 3, wherein said encoding the natural language problem corresponding to each of the SQL statements and generating the problem characterization vector corresponding to each of the SQL statements comprises:

inputting the natural language question corresponding to each SQL statement into a coding model for coding, and generating a question representation vector corresponding to each SQL statement;

the encoding the input question and generating the semantic representation vector of the input question comprise:

and inputting the input question into the coding model for coding to generate a semantic representation vector of the input question.

7. The method according to any one of claims 1-6, wherein the searching a problem characterization vector matching the semantic characterization vector in a search base according to the semantic characterization vector comprises:

8. A human-computer interaction device based on table data is characterized by comprising:

the system comprises an encoding module, a semantic representation module and a query processing module, wherein the encoding module is used for responding to a question-answer request based on table data, encoding an input question and generating a semantic representation vector of the input question;

and the feedback module is used for feeding back the answer information.

9. An electronic device, comprising: a processor, and a memory communicatively coupled to the processor;

the memory stores computer-executable instructions;

the processor executes computer-executable instructions stored by the memory to implement the method of any of claims 1-7.

10. A computer-readable storage medium having computer-executable instructions stored therein, which when executed by a processor, are configured to implement the method of any one of claims 1-7.