CN111522839A - Natural language query method based on deep learning - Google Patents

Natural language query method based on deep learning Download PDF

Info

Publication number
CN111522839A
CN111522839A CN202010336575.6A CN202010336575A CN111522839A CN 111522839 A CN111522839 A CN 111522839A CN 202010336575 A CN202010336575 A CN 202010336575A CN 111522839 A CN111522839 A CN 111522839A
Authority
CN
China
Prior art keywords
natural language
feature
query
vector
connection layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010336575.6A
Other languages
Chinese (zh)
Other versions
CN111522839B (en
Inventor
李玉华
李相臣
李瑞轩
辜希武
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN202010336575.6A priority Critical patent/CN111522839B/en
Publication of CN111522839A publication Critical patent/CN111522839A/en
Application granted granted Critical
Publication of CN111522839B publication Critical patent/CN111522839B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/242—Query formulation
    • G06F16/243—Natural language query formulation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/22—Indexing; Data structures therefor; Storage structures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/242—Query formulation
    • G06F16/2433—Query languages
    • G06F16/244—Grouping and aggregation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/242—Query formulation
    • G06F16/2433—Query languages
    • G06F16/2445—Data retrieval commands; View definitions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2452—Query translation
    • G06F16/24522—Translation of natural language queries to structured queries
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/28—Databases characterised by their database models, e.g. relational or object models
    • G06F16/284—Relational databases
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

The invention discloses a natural language query method based on deep learning, which comprises the following steps: s1, inputting the natural language question Q into a pre-trained sentence vector model to obtain a corresponding sentence vector; s2, retrieving the nearest neighbor sentence vector of the sentence vector corresponding to the natural language problem Q and the natural language problem thereof from the sentence vector space corresponding to the sentence vector model, and taking the relational database table of the natural language problem of the obtained nearest neighbor sentence vector as the target relational database table of the natural language problem Q; s3, splicing the natural language question Q and the header of the target relational database table together, and inputting the natural language question Q and the header into a conversion model which is trained well in advance to obtain a corresponding SQL query statement; and S4, querying in a target relational database table of the natural language question Q by using the obtained SQL statement to obtain a query result. The method can quickly determine the target relational database table after giving the natural language problem, and has higher accuracy of the query result in a real scene.

Description

Natural language query method based on deep learning
Technical Field
The invention belongs to the field of natural language processing, and particularly relates to a natural language query method based on deep learning.
Background
With the popularization and development of information systems, the quantity of data is more and more, and the importance is higher and more, and the demand of people for data is also increased. Sometimes, the data is too large and complex, which is a trouble for us, and it is not easy to find the needed data from the data, because the data is mostly stored in the database, and for the use of the database, a specific query language is needed to query the data from the database, and the learning and the use of the specific query language have certain thresholds, so that the learning cost is high for a common user, and there is an important meaning for researching a natural language query method.
Many researchers have proposed many natural language query methods in succession, and many attempts have been made in the direction of natural language query from early methods based on syntactic analysis to the latest methods based on deep learning. In the early days, scholars usually adopt grammar analysis to research a customized database natural language interface query system, the method usually needs domain expert knowledge and customization, and the generalization capability is general; with the rise of deep learning and the rapid promotion of computing power, researches on natural language query based on a deep learning method are gradually increased, and scholars successively put forward methods such as SEQ2SQL, SqlNet, Sqlova, TypeSQL, X-SQL, Coarse2Fine, Pointer-SQL, AntotatedSeq 2Seq, Multitask Question Answering Network (MQAN), Execution guided decoding and the like to process NL2SQL tasks, wherein the method has relatively good generalization capability effect, but needs a large amount of artificially labeled training data, adopts artificial labeling to generate the training data, and has relatively high manufacturing cost; meanwhile, some researchers propose to generate training data by using weak supervised Learning, such as memory enhanced Policy Optimization (MAPO) and metareward Learning (MeRL), and further perform deep Learning training by using the generated training data. However, in the above method, most target relational database tables of the current natural language problem are given, and only how to generate the SQL query statement is studied on the given tables, without considering the process of locating the target relational database tables of the natural language problem according to the natural language problem. In reality, the process of positioning the relational database table by the natural language problem is inevitable, and the relation is directly related to whether the correct SQL query statement can be generated, so that the accuracy of the query result in a real scene is low in the conventional method; most of the current research is concentrated in the field of English wikisq1 data sets, and the research on natural language query interfaces in the Chinese field is less.
Disclosure of Invention
Aiming at the defects or the improvement requirements of the prior art, the invention provides a natural language query method based on deep learning, which is used for solving the technical problem that the accuracy of query results in a real scene is low because the process of positioning a target relational database table of a natural language problem according to the natural language problem is not considered in the prior art.
In order to achieve the above object, in a first aspect, the present invention provides a deep learning-based natural language query method, including the following steps:
s1, inputting the natural language question Q into a pre-trained sentence vector model to obtain a corresponding sentence vector;
s2, retrieving the nearest neighbor sentence vector of the sentence vector corresponding to the natural language problem Q and the natural language problem thereof from the sentence vector space corresponding to the sentence vector model, and taking the relational database table of the natural language problem of the obtained nearest neighbor sentence vector as the target relational database table of the natural language problem Q;
s3, splicing the natural language question Q and the header of the target relational database table together, and inputting the natural language question Q and the header into a conversion model which is trained well in advance to obtain a corresponding SQL query statement;
s4, querying in a target relational database table of the natural language question Q by using the obtained SQL statement to obtain a query result;
wherein, the conversion model comprises: the device comprises a BERT pre-training model, a first full connection layer, a second full connection layer, a third full connection layer, a fourth full connection layer, a fifth full connection layer, a sixth full connection layer and a seventh full connection layer.
Further preferably, in step S3, the natural language question Q and the header of the target relational database table are spliced together to obtain an input list of the conversion model, specifically: [ [ CLS ], Q, [ SEP ], a first list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ], a second list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ],...., [ SEP ], a C list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ] ]; each character in the natural language question Q is an element of an input list, the position of each feature of each list head of the target relational database is a reserved position, and C is the number of columns of the database table.
Further preferably, a BERT pre-training model (Bidirectional Encoder Representation from transformations) is used to derive [ CLS ] based on the input list]Corresponding vector VCLSVector V corresponding to each character in natural language question QlAnd the vector V corresponding to each list head of the databasecAnd vectors corresponding to feature 1, feature 2, feature 3, and feature 4 of each list header are respectively denoted as Vc1、Vc2、Vc3And Vc4(ii) a And will VCLSRespectively inputting the V into a first full connection layer, a second full connection layer and a third full connection layerc1Inputting into a fourth fully-connected layer, and adding Vc2、Vc3And Vc4Respectively inputting the V into a fifth full connection layer and a sixth full connection layer in parallelc2、Vc3And Vc4And VlInputting the integrated data into a seventh full connection layer; l1, 2, theC, wherein C is the number of columns of the database table;
first full connection layer for [ CLS-based]Corresponding vector VCLSPredicting connectors among conditions of WHERE clauses in an SQL query statement;
second fully connected layer for [ CLS ] based]Corresponding vector VCLSPredicting the number of columns involved in the SELECT clause in the SQL query statement;
third fully connected layer for [ CLS ] based]Corresponding vector VCLSPredicting the number of columns involved in a WHERE clause in an SQL query statement;
the fourth full connection layer is used for generating a vector V corresponding to the characteristic 1 of each list headc1Predicting whether each column in the SQL query statement is a SELECT column and a corresponding aggregation symbol;
the fifth full connection layer is used for generating vectors V corresponding to the characteristics 2, 3 and 4 of each list headc2、Vc3、Vc4Predicting whether each column in the SQL query statement is a condition column in the WHERE clause;
the sixth full connection layer is used for generating a vector V corresponding to the feature 2, the feature 3 and the feature 4 of each list headc2、Vc3、Vc4Predicting comparison symbols of columns related to the WHERE clause in the SQL query statement;
the seventh full connection layer is used for solving the problem of the natural language Q according to the vector V corresponding to each characterlAnd a vector V corresponding to the feature 2, the feature 3, and the feature 4 of each list headerc2、Vc3、Vc4And predicting the starting position and the ending position of the value of each condition in the WHERE clause in the SQL query statement in the natural language question Q.
Further preferably, the first full connection layer, the second full connection layer, the third full connection layer, the fourth full connection layer, the fifth full connection layer, the sixth full connection layer, and the seventh full connection layer have the following sizes: hidden _ size x 3, hidden _ size x 4, hidden _ size x 7, hidden _ size x 1, hidden _ size x 4, and hidden _ size x 2; wherein hidden _ size is the dimension of the output vector of the BERT pre-training model.
Further preferably, the sentence vector model training method includes the following steps:
s011, collecting query data training set { (Q)i，SQLi) In which QiAnd SQLiThe ith natural language question and the corresponding SQL query statement are respectively; wherein, the respective natural language questions carry corresponding relational database tables; i is more than or equal to 1 and less than or equal to N, and N is the number of data in the query data training set;
s012, respectively calculating the similarity between every two SQL query sentences corresponding to the natural language problems in the query data training set to obtain a triple data set { (Q)i，Qj，Simij) In which SimijFor natural language question QiAnd QjThe similarity between the corresponding SQL query sentences, i is not equal to j; only considering the SELECT column and the WHERE conditional column when calculating the similarity of the SQL query statement;
s013, inputting the triple data set into the sentence vector model to train the triple data set to obtain a pre-trained sentence vector model, and inputting all natural language questions in the query data training set into the pre-trained sentence vector model to obtain a sentence vector space corresponding to the sentence vector model.
Further preferably, the training method of the conversion model includes:
training set { (Q) for query datai，SQLi) Splicing the respective language question with the corresponding database table header to obtain (Q)i，Headeri) (ii) a With (Q)i，Headeri) For input, SQLiAnd training the conversion model for outputting to obtain a pre-trained conversion model, wherein i is more than or equal to 1 and less than or equal to N, and N is the number of data in the query data training set.
Further preferably, the loss function of the conversion model is:
Figure BDA0002466616090000051
wherein los iss1Loss functions of connectors among all conditions of WHERE clauses in SQL query sentences in a query data training set and connectors among all conditions of WHERE clauses obtained by a first full connection layer of a conversion model are used; loss2A loss function based on the number of columns involved in a SELECT clause in an SQL query statement in a query data training set and the number of columns involved in the SELECT clause obtained by a second fully connected layer of the conversion model; loss3A loss function based on the number of columns involved in a WHERE clause in an SQL query statement in a query data training set and the number of columns involved in the WHERE clause obtained by a third full-connection layer of a conversion model; loss4Whether each column in the SQL query statement in the query data training set is a SELECT column, a corresponding aggregation symbol and a loss function of the SELECT column and the corresponding aggregation symbol is determined based on whether each column in the SQL query statement in the query data training set is the SELECT column and the corresponding aggregation symbol; loss5Whether each column in the SQL query statement in the query data training set is a condition column in a WHERE clause or not and a loss function of whether each column in the SQL query statement obtained by a fifth full-connection layer of the conversion model is a condition column in the WHERE clause or not are determined; loss6A loss function of the comparison symbol of the column involved in the WHERE clause in the SQL query sentence in the query data training set and the comparison symbol of the column involved in the WHERE clause obtained by the sixth full connection layer of the conversion model; loss7And the loss function is based on the starting and ending positions of the values of the conditions in the WHERE clauses in the SQL query sentences in the query data training set in the natural language problem and the starting and ending positions of the values of the conditions in the WHERE clauses obtained by the seventh full connection layer of the conversion model in the natural language problem.
In a second aspect, the present invention also provides a storage medium, which when read by a computer, causes the computer to execute the deep learning based natural language query method provided by the first aspect of the present invention.
Generally, by the above technical solution conceived by the present invention, the following beneficial effects can be obtained:
1. the invention provides a natural language query method based on deep learning, which is characterized in that a natural language question is input at will, a sentence vector is obtained through a sentence vector model, then a nearest neighbor sentence vector is searched in the obtained sentence vector space, a target relational database table of the nearest neighbor sentence vector is used as a target relational database table of the current natural language question, the natural language question is mapped to a vector space with fixed dimensionality through the steps, so that the distances between the sentence vectors corresponding to similar questions are closer in the vector space, and the database table where the nearest neighbor is located can be quickly found after the natural language question is given, namely the target relational database table of the natural language question, then, SQL sentences are generated according to the natural language question and then query is carried out on the obtained target relational database table, and the query result is obtained, and the accuracy of the query result in a real scene is higher.
2. According to the natural language query method based on deep learning, in the process of converting natural language into SQL query statements, a conversion model divides the whole SQL predicted task into different subtasks, and predicts the number of SELECT COLs, the columns and aggregators of the SELECT COLs, the number of condition columns in WHERE clauses, the columns of the condition columns, the comparison symbols of the condition columns, the values of the condition columns and the connectors between different conditions. All subtasks are combined and trained together, and finally, the results predicted by all subtasks are combined together, so that the model has high accuracy.
3. The natural language query method based on deep learning provided by the invention can quickly complete two processes of target database table positioning and target SQL query statement conversion for any natural language problem in actual scene application, has high processing speed and can meet the requirements of actual application.
4. The natural language query method based on deep learning provided by the invention has better mobility, and when the natural language query method is replaced to different fields, only the training data of a new field needs to be input into the model and retrained once.
5. The conversion model in the natural language query method based on deep learning provided by the invention adopts a subtask division method on the thought, and the specific conversion process is more refined, so that when the effect of the conversion model is evaluated, which subtask the short board of the conversion model is can be visually seen, further modification and promotion on the short board subtask are facilitated, and the accuracy of the model is higher.
Drawings
Fig. 1 is a flowchart of a natural language query method based on deep learning according to embodiment 1 of the present invention;
fig. 2 is a schematic structural diagram of a conversion model provided in embodiment 1 of the present invention;
fig. 3 is a flowchart of predicting connectors between conditions of a where clause by using a first full-link layer according to embodiment 1 of the present invention;
fig. 4 is a flowchart of predicting the number of columns in a select clause by using a second full link layer according to embodiment 1 of the present invention;
fig. 5 is a flowchart of predicting the number of conditional columns in a where clause by using a third full-link layer according to embodiment 1 of the present invention;
fig. 6 is a flowchart of predicting column and aggregation symbols in a select clause by using a fourth full link layer according to embodiment 1 of the present invention;
fig. 7 is a flowchart of predicting each condition column in a where clause by using a fifth full-link layer according to embodiment 1 of the present invention;
fig. 8 is a flowchart of comparing symbols in a where clause predicted by using a sixth full link layer according to embodiment 1 of the present invention;
fig. 9 is a flowchart for predicting the start and end positions of each condition value in a where clause by using a seventh full-link layer according to embodiment 1 of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
Examples 1,
A natural language query method based on deep learning, as shown in fig. 1, includes the following steps:
s1, inputting the natural language question Q into a pre-trained sentence vector model to obtain a corresponding sentence vector;
specifically, the sentence vector model training method comprises the following steps:
s011, collecting query data training set { (Q)i，SQLi) In which QiAnd SQLiThe ith natural language question and the corresponding SQL query statement are respectively; wherein, the respective natural language questions carry corresponding relational database tables; i is more than or equal to 1 and less than or equal to N, and N is the number of data in the query data training set;
s012, respectively calculating the similarity between every two SQL query sentences corresponding to the natural language problems in the query data training set to obtain a triple data set { (Q)i，Qj，Simij) In which SimijFor natural language question QiAnd QjThe similarity between the corresponding SQL query sentences, i is not equal to j; only considering the SELECT column and the WHERE conditional column when calculating the similarity of the SQL query statement;
specifically, [0, 1] is divided into 4 similarity intervals of the same size, which are [0, 0.25), [0.25, 0.5), [0.5, 0.75), [0.75, 1.0] on average, the obtained similarities are divided into the similarity intervals, and boundary values are taken down in the similarity intervals and mapped into a {0, 0.25, 0.5, 0.75, 1.0} set.
S013, inputting the triple data set into the sentence vector model to train the triple data set to obtain a pre-trained sentence vector model, and inputting all natural language questions in the query data training set into the pre-trained sentence vector model to obtain a sentence vector space corresponding to the sentence vector model.
Specifically, when the sentence vector model is used for calculating the sentence vector, the sentence vector model only comprises a BERT pre-training model and a pooling layer which are sequentially connected, and the output of the pooling layer is the sentence vector corresponding to the natural language problem. In the training process, the sentence vector model comprises a BERT pre-training model, a pooling layer, a splicing module and a soffmax layer which are connected in sequence; the loss function of the sentence vector model is a cross entropy loss function. Specifically, two problems are input into a BERT pre-training model, average pooling operation is carried out on output word vectors, then two sentence vectors u and v are spliced together in a [ u, v, | u-v | ] mode to obtain a vector with the dimension of 3 × n, wherein n is the dimension of the sentence vector, the obtained vector with the dimension of 3 × n is input into a softmax layer to obtain similarity categories of the two natural language problems, then a cross entropy loss function is adopted to calculate a loss value, and parameters in the sentence vector model are updated based on the obtained loss value.
S2, retrieving the nearest neighbor sentence vector of the sentence vector corresponding to the natural language problem Q and the natural language problem thereof from the sentence vector space corresponding to the sentence vector model, and taking the relational database table of the natural language problem of the obtained nearest neighbor sentence vector as the target relational database table of the natural language problem Q;
s3, splicing the natural language question Q and the header of the target relational database table to obtain an input list of the conversion model, which specifically comprises the following steps: [ [ CLS ], Q, [ SEP ], a first list header of a target relational database, a feature 1, a feature 2, a feature 3, a feature 4, [ SEP ], a second list header of the target relational database, a feature 1, a feature 2, a feature 3, a feature 4, [ SEP ], a C list header of the target relational database, a feature 1, a feature 2, a feature 3, a feature 4, [ SEP ] ], and inputting a person into a pre-trained conversion model to obtain a corresponding SQL query statement; each character in the natural language question Q is an element of an input list, the position of each feature of each list head of the target relational database is a reserved position, and C is the number of columns of the database table.
Specifically, as shown in fig. 2, the conversion model includes: the method comprises the following steps of (1) pre-training a BERT model, a first full-link layer, a second full-link layer, a third full-link layer, a fourth full-link layer, a fifth full-link layer, a sixth full-link layer and a seventh full-link layer;
wherein, in this embodiment, the size of first full tie layer, the full tie layer of second, the full tie layer of third, the full tie layer of fourth, the full tie layer of fifth, the full tie layer of sixth and the full tie layer of seventh is respectively: hidden _ size x 3, hidden _ size x 4, hidden _ size x 7, hidden _ size x 1, hidden _ size x 4, and hidden _ size x 2; wherein, hidden _ size is the dimension of output vector of the BERT pre-training model; before inputting the conversion model, splicing the natural language question Q and the table head of the target relational database table together to obtain an input list of the conversion model, wherein the method specifically comprises the following steps: [ [ CLS ], Q, [ SEP ], a first list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ], a second list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ],.. till., [ SEP ], a C list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ] ], wherein each character in the natural language question Q is an element of an input list, the position of each feature of each list header of the target relational database is a reserved position, the position of each feature of each list header of the target relational database is a word vector for obtaining the position output by a BERT pre-training model, and C is the number of columns of the database table.
The BERT pre-training model is used for obtaining [ CLS ] based on an input list]Corresponding vector VCLSVector V corresponding to each character in natural language question QlThe vector Vc corresponding to each list head of the database, and the vectors corresponding to the characteristic 1, the characteristic 2, the characteristic 3 and the characteristic 4 of each list head are respectively marked as Vc1、Vc2、Vc3 andVc4(ii) a And will VCLSRespectively inputting the V into a first full connection layer, a second full connection layer and a third full connection layerc1Inputting into a fourth fully-connected layer, and adding Vc2、Vc3And Vc4Respectively inputting the V into a fifth full connection layer and a sixth full connection layer in parallelc2、Vc3And Vc4And VlInputting the integrated data into a seventh full connection layer; l is 1, 2, theC is the column number of the database table; first full connection layer for [ CLS-based]Corresponding vector VCLSAnd predicting connectors between the conditions of the WHERE clause in the SQL query statement, wherein the connectors between the conditions of the WHERE clause in the embodiment include "and", "or" and no connector, as shown in fig. 3. Second fully connected layer for [ CLS ] based]Corresponding vector VCLSThe number of columns involved in the SELECT clause in the SQL query statement is predicted, as shown in fig. 4. Third fully connected layer for [ CLS ] based]Corresponding vector VCLSThe number of columns involved in the WHERE clause in the SQL query statement is predicted, as shown in fig. 5. The fourth full connection layer is used for generating a vector V corresponding to the characteristic 1 of each list headc1And predicting whether each column in the SQL query statement is a SELECT column and a corresponding aggregation symbol, as shown in fig. 6. The fifth full connection layer is used for generating vectors V corresponding to the characteristics 2, 3 and 4 of each list headc2、Vc3、Vc4And predicting whether each column in the SQL query statement is a conditional column in the WHERE clause, as shown in fig. 7. The sixth full connection layer is used for generating a vector V corresponding to the feature 2, the feature 3 and the feature 4 of each list headc2、Vc3、Vc4The comparison symbols of the columns involved in the WHERE clause in the SQL query statement are predicted, as shown in fig. 8. The seventh full connection layer is used for solving the problem of the natural language Q according to the vector V corresponding to each characterlAnd a vector V corresponding to the feature 2, the feature 3, and the feature 4 of each list headerc2、Vc3、Vc4And predicting the starting position and the ending position of the value of each condition in the WHERE clause in the SQL query statement in the natural language question Q, as shown in fig. 9.
Further, the training method of the conversion model comprises the following steps:
training set { (Q) for query datai，SQLi) Splicing the respective language question with the corresponding database table header to obtain (Q)i，Headeri) (ii) a With (Q)i，Headeri) For input, SQLiTraining the conversion model for output to obtain a pre-trained conversion model, wherein i is more than or equal to 1And N is less than or equal to N, wherein N is the number of data in the query data training set. Specifically, the method comprises the following steps:
s031, let i equal to 1;
s032 question Q of natural languageiSplicing the table head columns of the corresponding database table together to obtain an input list of the conversion model;
specifically, the input list is: [ [ CLS ]]，Qi，[SEP]First list header of target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP]Second list header of target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP]，......，[SEP]List head of C of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP]]Wherein, the natural language question QiEach character in (1) is an element of an input list, the position of each feature of each list head of the target relational database is a reserved position, and C is the number of columns of the database table.
S033, inputting the input list into a BERT pre-training model of the conversion model to obtain [ CLS]Corresponding vector VCLSVector V corresponding to each character in natural language question QlThe vector Vc corresponding to each list head of the database, and the vectors corresponding to the characteristic 1, the characteristic 2, the characteristic 3 and the characteristic 4 of each list head are respectively marked as Vc1、Vc2、Vc3 andVc4(ii) a L is 1, 2, the.
S034, mixing [ CLS]Corresponding vector VCLSInputting the data into a first full-connection layer of a conversion model, and predicting to obtain connectors among conditions of WHERE clauses in an SQL query statement;
s035, mixing [ CLS]Corresponding vector VCLSRespectively inputting the number of columns into a second full-connection layer and a third full-connection layer of the conversion model, and predicting to obtain the number of columns related to a SELECT clause and a WHERE clause in an SQL query statement;
s036, corresponding vector V to feature 1 of each list headc1Inputting the data into a fourth full-connection layer to predict whether each column in the SQL query statement is SELECT columns and corresponding aggregate symbols;
s037, corresponding vector V of feature 2, feature 3 and feature 4 of each list headc2Vc3Vc4Inputting the predicted results into a fifth full-connection layer and a sixth full-connection layer respectively, and predicting condition columns related to WHERE clauses in SQL query sentences and comparison signs of each condition column respectively;
s038, solving the natural language question QiThe vector corresponding to each word in (1) and the vector V corresponding to the feature 2, feature 3 and feature 4 of each list heada2、Vc3And Vc4Inputting the integrated result into a seventh full-connection layer, and predicting to obtain the starting position and the ending position of each condition value in the WHERE clause in the SQL query statement in the natural language question Q;
s039, preprocessing each part of elements of the SQL query statement of the training data into vectors which meet the specification size and serve as label, respectively solving loss functions with the results obtained in the steps S034-S038, adding each part of loss functions to serve as a total loss function, and updating parameter values in the conversion model through the result of the minimum term loss function;
specifically, processing each part of elements of the SQL query statement of the training data into a vector meeting the specification size as a label and solving for the loss function respectively needs the following processes:
(1) for subtask 1, the connection sign between conditions in the WHERE clause needs to be predicted, and the vector v with the size of 3 is predicted through the output of the first full-connection layer network13Respectively represent three connection symbols; converting specific connection symbols between WHERE clauses in SQL query statement into vectors v with the size of 111This subtask uses a cross entropy loss function:
loss1＝cross_entropy(v13，v11)
(2) for subtask 2, the number of columns involved in the SELECT clause needs to be predicted, predicted as a vector v of size 4 via the output of the second full-connection layer network24Represent different numbers of 1, 2, 3, 4 columns respectively; converting the specific value of the column number of the SELECT clause in the SQL query statement into a vector v with the size of 121This subtask uses a cross entropy loss function:
loss2＝cross_entropy(v24，v21)
(3) for subtask 3, the number of columns involved by the WHERE clause in the SQL query statement needs to be predicted, and the vector v with the size of 4 is predicted by the output of the third full-connection layer network34Represent different numbers of 1, 2, 3, 4 columns respectively; converting the specific value of the number of columns of the WHERE clause in the SQL query statement into a vector v with the size of 131This subtask uses a cross entropy loss function:
loss3＝cross_entropy(v34，v31)
(4) for the subtask 4, it is necessary to predict the column and aggregation symbol information in the SELECT clause in the SQL query statement, i.e., whether each column in the prediction header is a column in the SELECT clause and a specific aggregation symbol, and a vector v with a prediction size of 7 is output through the fourth full-connection-layer network47Respectively representing seven conditions that the current column is not a column in a SELECT clause, is a column obtained in the SELECT clause and has an aggregation symbol of "" AVG, MAX, MIN, SUM and COUNT "; converting the information of the SELECT clause in the SQL query statement into a vector v with the size of 141This subtask uses a cross entropy loss function:
loss4＝cross_entropy(v47，v41)
(5) for subtask 5, it is necessary to predict the column information in the WHERE clause, i.e. to predict whether each column in the header of the table is a column in the WHERE clause, and to output a vector v with a prediction size of 1 through the fifth full-connection-layer network51Representing the likelihood that the current column is a column in the WHERE clause; converting a particular column in a WHERE clause into a vector v of size 151', this subtask measures the error between the two using KL divergence:
loss5＝kl_loss(softmax(v51)，v51′)
(6) for subtask 6, the comparison sign of each condition column in the WHERE clause needs to be predicted, and a vector v with the prediction size of 4 is output through the sixth fully-connected network64Each representing 4 different ratiosComparing the signs; converting specific comparison symbols of each condition column in WHERE clause into vector v with size of 161This subtask uses a cross entropy loss function:
loss6＝cross_entropy(v64，v61)
(7) for subtask 7, it is necessary to predict the condition value of each condition in the WHERE clause, WHERE the start and end positions of each condition value in the natural language question are predicted, and output a vector v with a size of 3 × 2 × L via the seventh fully-connected network layer732LRepresenting possible starting and ending positions and probabilities of the current condition value, where L is the length of the natural language question; converting the specific starting and ending positions of the condition values in the WHERE clause in the natural language problem into a matrix v with the size of 3 x 2732This task uses a cross entropy loss function:
loss7＝cross_entropy(v732L，v732)
and synthesizing the subtasks to obtain a loss function of the conversion model as follows:
Figure BDA0002466616090000141
s040, where i is i + 1;
and S041, repeating the steps S032-S040 until i is larger than the number N of the data in the query data training set.
And S4, querying in a target relational database table of the natural language question Q by using the obtained SQL statement to obtain a query result.
Examples 2,
A storage medium that, when reading the instructions, causes a computer to execute the deep learning-based natural language query method provided in embodiment 1 of the present invention.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (8)

1. A natural language query method based on deep learning is characterized by comprising the following steps:
s1, inputting the natural language question Q into a pre-trained sentence vector model to obtain a corresponding sentence vector;
s2, retrieving the nearest neighbor sentence vector of the sentence vector corresponding to the natural language problem Q and the natural language problem thereof from the sentence vector space corresponding to the sentence vector model, and taking the relational database table of the natural language problem of the obtained nearest neighbor sentence vector as the target relational database table of the natural language problem Q;
s3, splicing the natural language question Q and the header of the target relational database table together, and inputting the natural language question Q and the header into a conversion model which is trained well in advance to obtain a corresponding SQL query statement;
s4, querying in a target relational database table of the natural language question Q by using the obtained SQL statement to obtain a query result;
wherein the conversion model comprises: the device comprises a BERT pre-training model, a first full connection layer, a second full connection layer, a third full connection layer, a fourth full connection layer, a fifth full connection layer, a sixth full connection layer and a seventh full connection layer.
2. The deep learning-based natural language query method according to claim 1, wherein in step S3, the natural language question Q and the header of the target relational database table are spliced together to obtain an input list of the conversion model, specifically: [ [ CLS ], Q, [ SEP ], a first list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ], a second list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ],...., [ SEP ], a C list header of the target relational database, feature 1, feature 2, feature 3, feature 4, [ SEP ] ]; each character in the natural language question Q is an element of an input list, the position of each feature of each list head of the target relational database is a reserved position, and C is the number of columns of the database table.
3. The deep learning-based natural language query method of claim 2, wherein the BERT pre-training model is used to derive [ CLS ] based on an input list]Corresponding vector VCLSVector V corresponding to each character in natural language question QlAnd the vector V corresponding to each list head of the databasecAnd vectors corresponding to feature 1, feature 2, feature 3, and feature 4 of each list header are respectively denoted as Vc1、Vc2、Vc3And Vc4(ii) a And will VCLSRespectively inputting the V into a first full connection layer, a second full connection layer and a third full connection layerc1Inputting into a fourth fully-connected layer, and adding Vc2、Vc3And Vc4Respectively inputting the V into a fifth full connection layer and a sixth full connection layer in parallelc2、Vc3And Vc4And VlInputting the integrated data into a seventh full connection layer; l is 1, 2, the.
The first full connection layer is used for being based on [ CLS]Corresponding vector VCLSPredicting connectors among conditions of WHERE clauses in an SQL query statement;
the second fully connected layer is used for being based on [ CLS]Corresponding vector VCLSPredicting the number of columns involved in the SELECT clause in the SQL query statement;
the third fully connected layer is used for being based on [ CLS]Corresponding vector VCLSPredicting the number of columns involved in a WHERE clause in an SQL query statement;
the fourth full connection layer is used for generating a vector V corresponding to the characteristic 1 of each list headc1Predicting whether each column in the SQL query statement is a SELECT column and a corresponding aggregation symbol;
the fifth full connection layer is used for generating a vector V corresponding to the characteristic 2, the characteristic 3 and the characteristic 4 of each list headc2、Vc3、Vc4Predicting whether each column in the SQL query statement is a condition column in the WHERE clause;
the sixth full connection layer is used for generating a vector V corresponding to the feature 2, the feature 3 and the feature 4 of each list headc2、Vc3、Vc4Predicting comparison symbols of columns related to the WHERE clause in the SQL query statement;
the seventh full connection layer is used for solving the problem that the vector V corresponding to each character in the natural language question Q islAnd a vector V corresponding to the feature 2, the feature 3, and the feature 4 of each list headerc2、Vc3、Vc4And predicting the starting position and the ending position of the value of each condition in the WHERE clause in the SQL query statement in the natural language question Q.
4. The deep learning-based natural language query method according to claim 1 or 3, wherein the first fully-connected layer, the second fully-connected layer, the third fully-connected layer, the fourth fully-connected layer, the fifth fully-connected layer, the sixth fully-connected layer and the seventh fully-connected layer have the sizes respectively: hidden _ size x 3, hidden _ size x 4, hidden _ size x 7, hidden _ size x 1, hidden _ size x 4, and hidden _ size x 2; wherein hidden _ size is the dimension of the output vector of the BERT pre-training model.
5. The deep learning-based natural language query method according to claim 1, wherein the sentence vector model training method comprises the following steps:
s011, collecting query data training set { (Q)i，SQLi) In which QiAnd SQLiThe ith natural language question and the corresponding SQL query statement are respectively; wherein, the respective natural language questions carry corresponding relational database tables; i is more than or equal to 1 and less than or equal to N, and N is the number of data in the query data training set;
s012, respectively calculating the similarity between every two SQL query sentences corresponding to the natural language problems in the query data training set to obtain a triple data set { (Q)i，Qj，Simij) In which SimijFor natural language question QiAnd QjCorresponding SQL query statementSimilarity between the two groups, i is not equal to j; only considering the SELECT column and the WHERE conditional column when calculating the similarity of the SQL query statement;
s013, inputting the triple data set into the sentence vector model to train the triple data set to obtain a pre-trained sentence vector model, and inputting all natural language questions in the query data training set into the pre-trained sentence vector model to obtain a sentence vector space corresponding to the sentence vector model.
6. The deep learning based natural language query method according to claim 1 or 3, wherein the training method of the conversion model comprises:
training set { (Q) for query datai，SQLi) Splicing the respective language question with the corresponding database table header to obtain (Q)i，Headeri) (ii) a With (Q)i，Headeri) For input, SQLiAnd training the conversion model for outputting to obtain a pre-trained conversion model, wherein i is more than or equal to 1 and less than or equal to N, and N is the number of data in the query data training set.
7. The deep learning based natural language query method of claim 6, wherein the loss function of the conversion model is:
Figure FDA0002466616080000041
therein, loss1Loss functions of connectors among all conditions of WHERE clauses in SQL query sentences in a query data training set and connectors among all conditions of WHERE clauses obtained by a first full connection layer of a conversion model are used; loss2A loss function based on the number of columns involved in a SELECT clause in an SQL query statement in a query data training set and the number of columns involved in the SELECT clause obtained by a second fully connected layer of the conversion model; loss3For WHERE clauses in SQL query sentences in training set based on query dataThe number of columns involved in (a) and the number of columns involved in the WHERE clause resulting from the third fully-connected layer of the conversion model; loss4Whether each column in the SQL query statement in the query data training set is a SELECT column, a corresponding aggregation symbol and a loss function of the SELECT column and the corresponding aggregation symbol is determined based on whether each column in the SQL query statement in the query data training set is the SELECT column and the corresponding aggregation symbol; loss5Whether each column in the SQL query statement in the query data training set is a condition column in a WHERE clause or not and a loss function of whether each column in the SQL query statement obtained by a fifth full-connection layer of the conversion model is a condition column in the WHERE clause or not are determined; loss6A loss function of the comparison symbol of the column involved in the WHERE clause in the SQL query sentence in the query data training set and the comparison symbol of the column involved in the WHERE clause obtained by the sixth full connection layer of the conversion model; loss7And the loss function is based on the starting and ending positions of the values of the conditions in the WHERE clauses in the SQL query sentences in the query data training set in the natural language problem and the starting and ending positions of the values of the conditions in the WHERE clauses obtained by the seventh full connection layer of the conversion model in the natural language problem.
8. A storage medium, which when read by a computer, causes the computer to execute the deep learning based natural language query method according to any one of claims 1 to 7.
CN202010336575.6A 2020-04-25 2020-04-25 Deep learning-based natural language query method Active CN111522839B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010336575.6A CN111522839B (en) 2020-04-25 2020-04-25 Deep learning-based natural language query method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010336575.6A CN111522839B (en) 2020-04-25 2020-04-25 Deep learning-based natural language query method

Publications (2)

Publication Number Publication Date
CN111522839A true CN111522839A (en) 2020-08-11
CN111522839B CN111522839B (en) 2023-09-01

Family

ID=71904930

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010336575.6A Active CN111522839B (en) 2020-04-25 2020-04-25 Deep learning-based natural language query method

Country Status (1)

Country Link
CN (1) CN111522839B (en)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112181420A (en) * 2020-08-27 2021-01-05 天津大学 Compiler defect positioning method based on reinforcement learning
CN112507098A (en) * 2020-12-18 2021-03-16 北京百度网讯科技有限公司 Question processing method, question processing device, electronic equipment, storage medium and program product
CN112559552A (en) * 2020-12-03 2021-03-26 北京百度网讯科技有限公司 Data pair generation method and device, electronic equipment and storage medium
CN112559690A (en) * 2020-12-21 2021-03-26 广东珠江智联信息科技股份有限公司 Natural language intelligent data modeling technology
CN112783921A (en) * 2021-01-26 2021-05-11 中国银联股份有限公司 Database operation method and device
CN113011136A (en) * 2021-04-02 2021-06-22 中国人民解放军国防科技大学 SQL (structured query language) analysis method and device based on correlation judgment and computer equipment
CN113282724A (en) * 2021-05-21 2021-08-20 北京京东振世信息技术有限公司 Interaction method and device for intelligent customer service
CN113743539A (en) * 2021-11-03 2021-12-03 南京云问网络技术有限公司 Form retrieval method based on deep learning
CN113761178A (en) * 2021-08-11 2021-12-07 北京三快在线科技有限公司 Data display method and device
CN114168619A (en) * 2022-02-09 2022-03-11 阿里巴巴达摩院(杭州)科技有限公司 Training method and device of language conversion model
CN116467347A (en) * 2023-03-22 2023-07-21 天云融创数据科技（北京）有限公司 Stock questioning and answering method

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090112835A1 (en) * 2007-10-24 2009-04-30 Marvin Elder Natural language database querying
US20110093271A1 (en) * 2005-01-24 2011-04-21 Bernard David E Multimodal natural language query system for processing and analyzing voice and proximity-based queries
CN106033466A (en) * 2015-03-20 2016-10-19 华为技术有限公司 Database query method and device
CN108717423A (en) * 2018-04-24 2018-10-30 南京航空航天大学 A kind of code segment recommendation method excavated based on deep semantic
CA3062071A1 (en) * 2017-05-18 2018-11-22 Salesforce.Com, Inc. Neural network based translation of natural language queries to database queries
WO2019081776A1 (en) * 2017-10-27 2019-05-02 Babylon Partners Limited A computer implemented determination method and system
CA3099828A1 (en) * 2018-06-27 2020-01-02 Bitdefender Ipr Management Ltd Systems and methods for translating natural language sentences into database queries
CN110727695A (en) * 2019-09-29 2020-01-24 浙江大学 Natural language query analysis method for novel power supply urban rail train data operation and maintenance
CN110888897A (en) * 2019-11-12 2020-03-17 杭州世平信息科技有限公司 Method and device for generating SQL (structured query language) statement according to natural language
CN110990536A (en) * 2019-12-06 2020-04-10 重庆邮电大学 CQL generation method based on BERT and knowledge graph perception

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110093271A1 (en) * 2005-01-24 2011-04-21 Bernard David E Multimodal natural language query system for processing and analyzing voice and proximity-based queries
US20090112835A1 (en) * 2007-10-24 2009-04-30 Marvin Elder Natural language database querying
CN106033466A (en) * 2015-03-20 2016-10-19 华为技术有限公司 Database query method and device
CA3062071A1 (en) * 2017-05-18 2018-11-22 Salesforce.Com, Inc. Neural network based translation of natural language queries to database queries
CN110945495A (en) * 2017-05-18 2020-03-31 易享信息技术有限公司 Conversion of natural language queries to database queries based on neural networks
WO2019081776A1 (en) * 2017-10-27 2019-05-02 Babylon Partners Limited A computer implemented determination method and system
CN108717423A (en) * 2018-04-24 2018-10-30 南京航空航天大学 A kind of code segment recommendation method excavated based on deep semantic
CA3099828A1 (en) * 2018-06-27 2020-01-02 Bitdefender Ipr Management Ltd Systems and methods for translating natural language sentences into database queries
CN110727695A (en) * 2019-09-29 2020-01-24 浙江大学 Natural language query analysis method for novel power supply urban rail train data operation and maintenance
CN110888897A (en) * 2019-11-12 2020-03-17 杭州世平信息科技有限公司 Method and device for generating SQL (structured query language) statement according to natural language
CN110990536A (en) * 2019-12-06 2020-04-10 重庆邮电大学 CQL generation method based on BERT and knowledge graph perception

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
XIAOYU ZHANG, FENGJING YIN , GUOJIE MA, BIN GE, AND WEIDONG XIAO: "《M-SQL Multi-Task Representation Learning for Single-Table Text2sql Generation》", vol. 8, pages 43159 - 43167 *
宋彦京: "《基于MVC架构的智能认知系统的设计与实现》", 《信息科技》 *

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112181420A (en) * 2020-08-27 2021-01-05 天津大学 Compiler defect positioning method based on reinforcement learning
CN112181420B (en) * 2020-08-27 2022-06-28 天津大学 Compiler defect positioning method based on reinforcement learning
CN112559552A (en) * 2020-12-03 2021-03-26 北京百度网讯科技有限公司 Data pair generation method and device, electronic equipment and storage medium
US11748340B2 (en) 2020-12-03 2023-09-05 Beijing Baidu Netcom Science And Technology Co., Ltd. Data pair generating method, apparatus, electronic device and storage medium
CN112559552B (en) * 2020-12-03 2023-07-25 北京百度网讯科技有限公司 Data pair generation method and device, electronic equipment and storage medium
CN112507098B (en) * 2020-12-18 2022-01-28 北京百度网讯科技有限公司 Question processing method, question processing device, electronic equipment, storage medium and program product
CN112507098A (en) * 2020-12-18 2021-03-16 北京百度网讯科技有限公司 Question processing method, question processing device, electronic equipment, storage medium and program product
CN112559690A (en) * 2020-12-21 2021-03-26 广东珠江智联信息科技股份有限公司 Natural language intelligent data modeling technology
CN112783921A (en) * 2021-01-26 2021-05-11 中国银联股份有限公司 Database operation method and device
CN113011136A (en) * 2021-04-02 2021-06-22 中国人民解放军国防科技大学 SQL (structured query language) analysis method and device based on correlation judgment and computer equipment
CN113282724A (en) * 2021-05-21 2021-08-20 北京京东振世信息技术有限公司 Interaction method and device for intelligent customer service
CN113761178A (en) * 2021-08-11 2021-12-07 北京三快在线科技有限公司 Data display method and device
CN113743539A (en) * 2021-11-03 2021-12-03 南京云问网络技术有限公司 Form retrieval method based on deep learning
CN114168619A (en) * 2022-02-09 2022-03-11 阿里巴巴达摩院(杭州)科技有限公司 Training method and device of language conversion model
CN116467347A (en) * 2023-03-22 2023-07-21 天云融创数据科技（北京）有限公司 Stock questioning and answering method

Also Published As

Publication number Publication date
CN111522839B (en) 2023-09-01

Similar Documents

Publication Publication Date Title
CN111522839A (en) Natural language query method based on deep learning
CN113239700A (en) Text semantic matching device, system, method and storage medium for improving BERT
CN111666427B (en) Entity relationship joint extraction method, device, equipment and medium
CN110134946B (en) Machine reading understanding method for complex data
CN110309511B (en) Shared representation-based multitask language analysis system and method
CN112199532B (en) Zero sample image retrieval method and device based on Hash coding and graph attention machine mechanism
CN112163429B (en) Sentence correlation obtaining method, system and medium combining cyclic network and BERT
CN113743119B (en) Chinese named entity recognition module, method and device and electronic equipment
CN115858788A (en) Visual angle level text emotion classification system based on double-graph convolutional neural network
CN111241209A (en) Method and apparatus for generating information
CN114638228A (en) Chinese named entity recognition method based on word set self-attention
CN113449489B (en) Punctuation mark labeling method, punctuation mark labeling device, computer equipment and storage medium
Lin et al. Research on named entity recognition method of metro on-board equipment based on multiheaded self-attention mechanism and CNN-BiLSTM-CRF
CN112905750A (en) Generation method and device of optimization model
CN113392191B (en) Text matching method and device based on multi-dimensional semantic joint learning
CN113792120B (en) Graph network construction method and device, reading and understanding method and device
CN112487811B (en) Cascading information extraction system and method based on reinforcement learning
CN114842301A (en) Semi-supervised training method of image annotation model
CN113901813A (en) Event extraction method based on topic features and implicit sentence structure
CN113051920A (en) Named entity recognition method and device, computer equipment and storage medium
CN112560487A (en) Entity relationship extraction method and system based on domestic equipment
CN112052685A (en) End-to-end text entity relationship identification method based on two-dimensional time sequence network
CN114626378A (en) Named entity recognition method and device, electronic equipment and computer readable storage medium
CN115408506B (en) NL2SQL method combining semantic analysis and semantic component matching
CN114548325B (en) Zero sample relation extraction method and system based on dual contrast learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant