KR20230081594A - Apparatus and method for processing natural language query about relational database using transformer neural network - Google Patents

Abstract

A natural language query processing apparatus according to an embodiment for achieving the object of the present invention receives a natural language query input by a user, generates a structured query based on the natural language query, and a processor performs a structured query based on the natural language query. In generating a query, a structured query is created using a result of cross-attention generated between a schema relationship extracted based on a relationship between a natural language processing result of a natural language query and a lower database in a database related to the natural language query.

Description

APPARATUS AND METHOD FOR PROCESSING NATURAL LANGUAGE QUERY ABOUT RELATIONAL DATABASE USING TRANSFORMER NEURAL NETWORK}

The present invention relates to a query processing method and apparatus for a database, and relates to a technology for generating a standardized/structured query based on a natural language query. In particular, the present invention relates to a technique for generating a standardized/structured query based on a natural language query using a transformer neural network and a relational database.

With the recent spread of the Internet, users are exposed to a vast amount of information. In order to acquire only the necessary information among such vast amounts of information, an effective search is required.

A search is made in response to a query. A typical question answering system presents an answer as a result of a question. It is very difficult to query and search exactly what information requesters want.

In order to obtain the desired result, it is required that the query be a structured query with specifications. However, in order to create a structured query, specialized knowledge of specifications/rules is sometimes required, so most users find it difficult to create a structured query on their own.

For this reason, there have been attempts by users to generate queries in natural language and interpret/convert the natural language queries into standardized/structured queries.

Korean Patent Registration No. KR 10-2150908 "Natural Language Query Interpretation Method and System" proposes a query processing method that more clearly interprets the user's intention and effectively meets the user's intention in interpreting the natural language query.

However, even with these prior arts, the process of converting a natural language query into a structured query is not perfect, and an effective technique for converting a natural language query into a structured query is required.

Korean Registered Patent No. KR 10-2150908 "Natural language query interpretation method and system" (published on August 27, 2020) Korean Registered Patent No. KR 10-2277787 "Method for Predicting Columns and Tables Used in Translating SQL Queries from Natural Language Based on Neural Networks" (Published on July 9, 2021) Korean Registered Patent No. KR 10-2345568 "Method for connecting natural language words with database columns and tables" (published on December 27, 2021)

"Attention Is All You Need", Ashish Vaswani et al., 31st Conference on Neural Information Processing Systems (NIPS 2017), (published on December 6, 2017)

The object of the present invention to solve the above problems is a search system that is expected to be applied not only to database management systems but also to future search systems, and to find related data or correct answers when a user sends a keyword or query to desired information. As a technology, an apparatus and method for encoding a transformer neural network-based natural language query and a relational database into a meaningful representation that can be utilized when using unstructured data as well as structured data such as a database are provided.

An object of the present invention is to provide a query response result desired by a user by automatically generating a structured query even when a user inputs a natural language query without knowledge of the structured query.

An object of the present invention is to generate a structured query that meets the user's intention and provide a query response result by pre-analyzing and encoding schema information that can be obtained only with a natural language query input by a user.

An object of the present invention is to advance a process of generating a structured query and query response result when a user inputs a natural language query.

An object of the present invention is to propose a new neural network structure based on cross-attention that advances the process of generating a structured query in response to a natural language query input.

An object of the present invention is to propose a scheme for extracting schema relationships based on relationships between sub-databases, which advances the process of generating a structured query in response to a natural language query input.

A natural language query processing apparatus according to an embodiment for achieving the object of the present invention includes a processor. The processor receives a natural language query input by a user, generates a structured query based on the natural language query, and in generating the structured query based on the natural language query, the natural language processing result for the natural language query and the natural language query A structured query is created using the mutually generated cross-attention result of schema relationships extracted based on relationships between sub-databases in the database related to .

The processor may generate a structured query using a first cross-attention result indicating a degree of relevance of a natural language processing result with a schema relationship and a second cross-attention result indicating a degree of relevance of a schema relationship with a natural language processing result, The cross-attention result may include a first cross-attention result and a second cross-attention result.

A natural language query processing apparatus according to an embodiment of the present invention includes a natural language processing model that generates a natural language processing result based on a natural language query; and a schema encoder model outputting a schema relationship.

The schema encoder model may output a schema semantic expression as a schema relationship based on an input schema graph.

The natural language processing model may receive a preprocessing result of a natural language query and output a semantic expression of a natural language element as a natural language processing result.

A natural language query processing apparatus according to an embodiment of the present invention includes a first cross-attention layer generating a first cross-attention result indicating a degree of relevance of a result of natural language processing with a schema relationship; and a second cross-attention layer generating a second cross-attention result indicating a degree of relevance of the schema relationship with the natural language processing result.

The natural language processing model may include a first inner attention layer having a self-attention function therein, and the schema encoder model may include a second inner attention layer having at least one function of self-attention and cross-attention therein. .

The processor may block updating of parameters within the natural language processing model based on the output of the first inner attention layer.

The processor may block updating of parameters within the natural language processing model based on the cross-attention result.

The apparatus for processing a natural language query according to an embodiment of the present invention may further include a decoder layer generating a structured query based on a cross attention result.

A natural language query processing method according to an embodiment of the present invention may be executed by a computing system including a processor. A natural language query processing method according to an embodiment of the present invention includes receiving a natural language query input by a user; and generating a structured query based on the natural language query, wherein generating the structured query comprises: obtaining a natural language processing result for the natural language query; obtaining a schema relationship extracted based on a relationship between sub-databases in a database related to a natural language query; obtaining a cross-attention result generated between the natural language processing result and the schema relationship; and generating a structured query using the cross attention result.

The obtaining of the cross-attention result may include: obtaining a first cross-attention result indicating a degree of relation between the natural language processing result and the schema relationship; and obtaining a second cross-attention result indicating a degree of relevance of the schema relationship with the natural language processing result.

In the step of obtaining a schema relationship, a schema semantic expression may be obtained as a schema relationship based on an input schema graph. In the step of obtaining a natural language processing result, a preprocessing result of a natural language query may be input and a semantic expression of a natural language element may be obtained as a natural language processing result.

Obtaining the natural language processing result may be performed in a state in which updating of internal parameters of the natural language processing model based on the output of the internal attention layer is blocked.

Obtaining the natural language processing result may be performed in a state in which updating of internal parameters of the natural language processing model based on the cross attention result is blocked.

A natural language query processing method executed by a computing system including a processor according to an embodiment of the present invention provides a one-to-one correspondence between sub-databases in a database related to a natural language query, and extracting as either one-to-many correspondence; and generating a structured query from the natural language query based on the schema relationship.

The natural language query processing method according to an embodiment of the present invention may further include predicting a schema relationship as one of a mandatory relationship and an optional relationship through comparison between lower databases.

A natural language query processing method according to an embodiment of the present invention includes encoding a schema relationship using a first natural language model corresponding to a first sub-database and a second natural language model corresponding to a second sub-database among sub-databases. can include more.

The natural language query processing method according to an embodiment of the present invention may further include pre-encoding at least a portion of a schema relationship by using mask column modeling for a portion of lower databases.

The natural language query processing method according to an embodiment of the present invention may further include obtaining a natural language processing result based on the natural language query. In this case, the generating of the structured query may include obtaining a result of natural language processing and a cross-attention result generated between schema relationships; and generating a structured query using the cross attention result.

The natural language query processing method according to an embodiment of the present invention may further include obtaining a natural language processing result based on a natural language query using a pre-learned language model. In this case, the step of obtaining the natural language processing result may be performed in a state in which updating of internal parameters of the language model is blocked.

According to an embodiment of the present invention, in translating a natural language query into a standardized/structured query such as an SQL query, it is possible to more accurately calculate a semantic expression for a natural language query and a database schema, thereby increasing the performance of the translation system.

According to an embodiment of the present invention, even when a user inputs a natural language query without knowledge of the structured query, a structured query may be automatically generated and a query response result desired by the user may be provided.

According to an embodiment of the present invention, it is possible to generate a structured query that meets the user's intention and provide a query response result by pre-analyzing and encoding schema information that can be obtained only with a natural language query input by a user.

According to an embodiment of the present invention, when a user inputs a natural language query, a process of generating a structured query and query response result can be advanced.

According to an embodiment of the present invention, it is possible to implement a new neural network structure based on cross attention that advances the process of generating a structured query in response to a natural language query input.

According to an embodiment of the present invention, a schema relationship based on a relationship between sub-databases may be extracted, which enhances a process of generating a structured query in response to a natural language query input.

1 is a conceptual diagram illustrating a system including a natural language query processing apparatus according to an embodiment of the present invention.
2 is a conceptual diagram illustrating an example of a natural language processing scenario executed by a natural language query processing apparatus according to an embodiment of the present invention.
3 is a conceptual diagram illustrating a process of generating an equivalent query by a virtual user to explain a process of processing a natural language query executed by a natural language query processing apparatus according to an embodiment of the present invention.
4 is a conceptual diagram illustrating a language model, a schema encoder, and a cross-attention layer for natural language processing as an internal structure of a natural language query processing apparatus according to an embodiment of the present invention.
5 is a conceptual diagram illustrating an internal structure of a natural language query processing apparatus according to an embodiment of the present invention, when updating of internal parameters of a language model is blocked.
6 is a conceptual diagram illustrating an embodiment further including a decoder layer generating a structured query based on a result of cross attention as an internal structure of a natural language query processing apparatus according to an embodiment of the present invention.
7 is an operational flowchart illustrating a natural language query processing method executed by a natural language query processing apparatus according to an embodiment of the present invention.
FIG. 8 is an operation flow chart showing steps S350 of FIG. 7 and the processes before and after it in more detail.
9 is an operational flowchart illustrating a natural language query processing method according to another embodiment of the present invention.
10 is a conceptual diagram illustrating an embodiment of a schema linking process when a natural language query and a schema related to automobiles are given in a natural language query processing method performed by a natural language query processing apparatus according to an embodiment of the present invention.
11 illustrates that in a natural language query processing method performed by a natural language query processing apparatus according to an embodiment of the present invention, one natural language query can be translated into different structured queries according to relationships between sub-databases (tables) in a database. It is a conceptual diagram to explain.
12 is a conceptual diagram illustrating an embodiment of extracting relationships between lower databases (tables) by utilizing database record values in a natural language query processing method performed by a natural language query processing apparatus according to an embodiment of the present invention.
13 is a conceptual diagram illustrating an example of schema encoding in a natural language query processing method performed by a natural language query processing apparatus according to an embodiment of the present invention.
14 is a block diagram illustrating a generalized natural language processing device or computing system according to an embodiment of the present invention.

Since the present invention can have various changes and various embodiments, it will be described in detail by exemplifying specific embodiments in the drawings. However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention. Like reference numerals have been used for like elements throughout the description of each figure.

Terms such as first, second, A, and B may be used to describe various components, but the components should not be limited by the terms. These terms are only used for the purpose of distinguishing one component from another. For example, a first element may be termed a second element, and similarly, a second element may be termed a first element, without departing from the scope of the present invention. The term “and/or” includes any combination of a plurality of related listed items or any of a plurality of related listed items.

It is understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, but other elements may exist in the middle. It should be. On the other hand, when an element is referred to as “directly connected” or “directly connected” to another element, it should be understood that no other element exists in the middle.

Terms used in this application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, the terms "include" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and unless explicitly defined in the present application, they should not be interpreted in an ideal or excessively formal meaning. don't

Korean Patent Registration No. KR 10-2150908 "Natural Language Query Interpretation Method and System", Korean Patent Registration No. KR 10-2277787 "Method for Predicting Columns and Tables Used in SQL Query Translation from Neural Network-based Natural Language", A method for translating a natural language query into a structured query disclosed in Korean Patent Registration No. KR 10-2345568 "Method of connecting natural language words with columns and tables of a database", etc. A method for analyzing the relationship between databases and prior studies, including transformer neural network structure disclosed in "Attention Is All You Need", Ashish Vaswani et al., 31st Conference on Neural Information Processing Systems (NIPS 2017) (December 6, 2017) Matters disclosed in the literature may be included as part or all of the configuration of the present invention within the scope consistent with the purpose of the present invention. Since those skilled in the art will be able to clearly infer the connection between the object and configuration of the present invention from the contents of the prior art documents, excessively detailed descriptions that may obscure the purpose of the present invention will be omitted, and the prior art documents will be introduced. substitute for

Hereinafter, with reference to the accompanying drawings, preferred embodiments of the present invention will be described in more detail. In order to facilitate overall understanding in the description of the present invention, the same reference numerals are used for the same components in the drawings, and redundant descriptions of the same components are omitted.

1 is a conceptual diagram illustrating a system including a natural language query processing apparatus according to an embodiment of the present invention.

Referring to Figure 1, the system can be understood as a system for translating natural language queries to a database. Also, the system may be understood as a system that automatically generates a response result corresponding to the natural language query when the user 100 inputs a natural language query to the database.

The query translator 240 in the system may include a transformer neural network therein, and may encode a natural language query and a schema extraction result of a relational database into a meaning representation based on the transformer neural network. The query translator 240 may decode the encoded result, convert it into a structured query, and transmit it to the answer engine 260 .

In order to use a general database management system (DBMS), a general user must understand the structure of the database and the syntax for creating a structured query. At this time, general users need knowledge of the schema according to the structure of the database in order to properly write and input a structured query. At this time, the structured query may be a generally well-known Structured Query Language (SQL) or may be a structured query of another known type. An example of schema knowledge and structured query knowledge required by a general user will be described in FIG. 3 to be described later.

In the system of FIG. 1, user 100 may not require direct knowledge of structured queries and/or schemas. The user 100 may input a natural language query into the conversation manager 220 . In this case, the natural language query input method may be text input via a keyboard/keypad or voice input. In the case of voice input, a pre-processing process of converting a voice recognition result into a natural language query may be performed by the dialog manager 220 .

The dialog manager 220 may transmit a query formed in natural language to the query translator 240 . Query translator 240 may translate a natural language query input into a structured query. Query translator 240 may pass the structured query to response engine 260 .

The response engine 260 may include functions such as a known artificial intelligence-based interactive assistant. Response engine 260 may provide search results in the database corresponding to the structured query. The response engine 260 may provide the search results to the user 100 as natural language responses and/or visualized results.

Although not shown in FIG. 1 for convenience of description, the dialog manager 220, query translator 240, and response engine 260 of FIG. 1 may be electronically connected to at least one processor and managed and controlled by the processor. there is. The processor is electronically connected to at least one memory, and at least one command stored in the memory is executed by the processor, thereby providing functions of the dialog manager 220, the query translator 240, and the response engine 260 of FIG. 1. It can be. In another embodiment of the present invention, at least some of the dialog manager 220, query translator 240, and response engine 260 of FIG. 1 may be implemented in the form of dedicated hardware.

A system/device according to an embodiment for achieving the object of the present invention may include a function of encoding a transformer neural network-based natural language query and a relational database into a meaningful expression.

In a situation where multilingual speech recognition is required, the system/device performs multilingual speech recognition with at least one artificial intelligence model when only a single language can be recognized due to hardware limitations of the terminal device, and automatically recognizes and processes the user's utterance. Voice recognition service can be provided.

The system/device passes audio input in English (EN) and Korean (KO) through a 7-layer Convolutional Neural Network (CNN) feature extractor, and then extracts voice feature information through a 24-layer Transformer Encoder, The final voice recognition result can be output through the CTC Projection output layer capable of outputting Korean (KO), English (EN), or both Korean and English (ALL).

In the system/device, the output unit of each CTC projection output layer is a Unicode unit corresponding to 1 byte (8 bits), and it can be recombined to generate a final result. At this time, the information in byte units may be expressed as 0x00 to 0xff in an octal representation method. Unicode representation in byte units can express all languages in the world as a combination of 256 (2 to the 8th power) of 8 bits (1 byte), so even if the number of supported languages increases, the size of the entire model supports 256 levels of output. It can only increase as much as the CTC Projection output layer.

The system/device may be implemented such that all recognized languages are commonly shared and have a separate CTC projection output layer for each language. If the speech language classification is not certain, a universal (ALL) output layer in which all languages are simultaneously learned and all languages can be output may be used. The general-purpose output layer performs worse than the dedicated (KO or EN) output layer, but it can prevent malfunctions that generate output of the wrong language when speech language recognition is not performed properly.

The system/device may be equipped with a speech language classification model to improve the performance of multilingual speech recognition.

The voice language classification model of the system/device is based on the Wav2Vec2.0 model, divides voice information into 25ms (millisecond) units, extracts voice information using the Wav2Vec2.0 model, and extracts 25ms units through the Projector layer. The average value of the received speech information (mean pooling operation performed) can be classified through a speech language classifier.

When the system/device inputs the audio data received as input to the Wav2Byte speech recognizer and the Wav2Vec2.0-based speech language classifier, respectively, and selects the CTC projection output layer for each language of the speech recognizer of the Wav2Byte structure, the Wav2Vec2.0-based speech language classifier The final speech recognition result can be generated by selecting an output layer using the output (language classification information). When the confidence of speech language classification is less than 70%, the system/device may generate a speech recognition result using an ALL language output layer instead of a dedicated language output layer in order to prevent malfunction.

2 is a conceptual diagram illustrating an example of a natural language processing scenario executed by a natural language query processing apparatus according to an embodiment of the present invention.

Referring to FIG. 2 , a user 100 uses a database management system (DBMS) 200 and a conversational interface via the dialog manager 220, the query translator 240, and the response engine 260 of FIG. 1 A scenario in which queries and search results are provided is shown.

The user 100 may input a natural language query for a city having an ATM transaction volume exceeding 300,000. The DBMS 200 may report that a city with a transaction volume exceeding 300,000 is not searched, but that Seoul is searched as a city having 299,580 transactions closest to 300,000. In FIG. 1, it is assumed that the user 100 is a bank or financial institution employee, but the scope of the present invention is not limited to this assumption.

At this time, referring to FIGS. 1 and 2 together, at least one of the query translator 240 and the response engine 260 of FIG. 1 cooperates with each other or independently performs auxiliary structuring based on the search result according to the structured query. The query may be additionally generated, and an intelligent response based thereon may be provided to the user 100. For example, if an appropriate number of results are retrieved as a query search result, the original search result may be provided to the user 100 as it is, and if an appropriate number of results are not retrieved, an auxiliary search result may be provided to the user 100 together. .

The user 100 may input a second query after receiving the first query response result. At this time, the user 100 may input a natural language query while partially omitting what the search is for, and the DBMS 200 generates a structured query that meets the user 100's intention based on the schema relationship, A query may be retrieved from the database, and a query response result may be provided to the user 100 .

3 is a conceptual diagram illustrating a process of generating an equivalent query by a virtual user to explain a process of processing a natural language query executed by a natural language query processing apparatus according to an embodiment of the present invention.

Referring to FIG. 3 , a user 100 may input a natural language query to the natural language query processing apparatus according to an embodiment of the present invention. In this case, a virtual user 120 is assumed to explain the internal operation of the natural language query processing device. Virtual user 120 may generate a structured query in response to user 100's natural language query.

The structured query generated by the virtual user 120 may be a query equivalent to the natural language query of the user 100 in the system of FIG. 3 .

In other words, in the prior art, the virtual user 120 needs to directly write and input a structured query as shown in FIG. 3 . At this time, the virtual user 120 can know that in order to identify a city where more than 300,000 transactions have occurred, the corresponding user must be familiar with three table schemas and must be able to write a complex nested SQL query.

However, the general public does not know about the SQL language, and it is not easy for even experts with knowledge of SQL queries to write queries that accurately reflect their intentions, taking into account the uniqueness of databases.

In addition, it is difficult for users of the prior art DBMS to check whether they have correctly written SQL queries.

For example, even if the virtual user 120 correctly writes the SQL-based query, if the query response result returned by the DBMS is provided as no search result, the user may not be sure that the user 120 correctly writes the SQL query.

However, in the embodiment of FIG. 3, the user 100 can query the DBMS 200 in natural language, the user 200 can accurately express desired information in words, and understand the schema of a complex sub-database (table). no need.

In one embodiment of the present invention, a neural network-based natural language processing technique may be used. Such natural language processing technology may use a known transformer neural network or the like, and may be included as part of the configuration of the present invention, if necessary.

In recent years, several natural language query translation technologies based on neural networks have emerged. In the early days, studies based on recurrent neural networks/Long Short Term Memory networks were introduced, but when the input length became long, gradient exploding/vanishing problems occurred. Structured neural networks are widely used.

RAT-SQL is a representative technology that utilizes the transformer structure. RAT-SQL is an encoder-decoder model that simultaneously puts a natural language query and schema into a pre-learned language model BERT to calculate a meaning representation, and uses a relational recognition transformer neural network to express explicit information between input tokens. Calculate the final semantic expression of tokens using relational information.

In order to input a natural language query and schema into a transformer neural network of the prior art, schema information must be formed into continuous words and undergo a tokenization process. RAT-SQL lists the table names and column names included in the schema in a row and connects them with a natural language query with a special token called [SEP] (separator) in between.

In addition, in order to use a relation-aware transformer, an explicit relationship between tokens must be extracted. The relationship between schema tokens uses a connection relationship on the schema graph, and the relationship between a natural language query and schema tokens is schema linking (schema -linking) to find out. With the final semantic expression calculated in this way, the decoder sequentially generates SQL phrases.

Although various natural language query translation technologies have been added since RAT-SQL, most of them follow the RAT-SQL encoder method. A method of effectively initializing model parameters in an encoding method was proposed, and then a decoder was newly proposed.

The RAT-SQL model can greatly increase the translation performance of natural language queries by utilizing the pretrained language model BERT.

However, the prior art RAT-SQL method using BERT has several problems.

First, BERT is trained to receive natural language sentences, and RAT-SQL receives input by enumerating table and column names together with natural language query sentences. In this case, BERT needs additional training to extract effective semantic expressions for new input formats, but there is not much training data for natural language queries and schema pairs.

Second, BERT has a disadvantage that the input length is limited to 512 tokens, so the current structure cannot be applied to a schema with a large number of columns and tables.

Thirdly, there is a disadvantage that the relationship information of input tokens cannot be used effectively.

RAT-SQL, which is a prior art, can explicitly input a connection relationship between tokens to a relation recognition transformer in order to utilize a connection relationship between tokens including columns and tables.

However, the relationship-aware transformer is a module used at the back of the encoder, and there is a problem in that the relationship information between tokens is not utilized at all in the front part where the input token is first given. Such a model structure is likely to affect the later part of the model by creating incorrect intermediate semantic representations.

Another prior art uses meta data or schema to enhance the process of interpreting a natural language query. For example, in Korean Registered Patent No. KR 10-2150908 "Natural Language Query Interpretation Method and System" described above, the first types of each natural language term determined based on morpheme analysis are further detailed second type/subtypes using a schema. Configurations embodied in component patterns are disclosed.

However, although these conventional techniques can extract the meaning of each term constituting a natural language query in more detail, they still cannot extract all information required to generate a structured query (eg, SQL query).

In order to solve the problems of the prior art, the present invention proposes a new internal structure capable of upgrading a natural language query processing apparatus.

4 is a conceptual diagram illustrating a language model, a schema encoder, and a cross-attention layer for natural language processing as an internal structure of a natural language query processing apparatus according to an embodiment of the present invention.

4 illustrates a process of encoding a natural language query and a schema of an apparatus for encoding a transformer neural network-based natural language query and a relational database into a meaningful expression according to an embodiment of the present invention, and then recalculating a semantic expression in consideration of the two relationships. It is a concept for Although not shown in FIG. 4 , each component of FIG. 4 may be electronically connected to a processor, and operations and functions of each component may be controlled and managed by the processor.

Referring to FIG. 4 , the natural language query processing apparatus according to an embodiment of the present invention is a query translator 240 that receives a natural language query input by a user 100 and generates a structured query based on the natural language query. can

In generating a structured query based on a natural language query, a natural language query processing apparatus according to an embodiment of the present invention generates a structured query based on 1) a natural language query result based on a relationship between a natural language query result and a sub-database in a database related to the natural language query. A structured query can be created using the result of 2) cross attention generated between extracted schema relationships.

A natural language query processing apparatus according to an embodiment of the present invention includes a first cross-attention result indicating the degree of relevance of a natural language processing result to a schema relationship, and a second cross-attention result indicating a degree of relevance of a schema relationship to a natural language processing result. A structured query may be created using the query, and the cross-attention result may include a first cross-attention result and a second cross-attention result.

A natural language query processing apparatus according to an embodiment of the present invention includes a language model 244 as a natural language processing model that generates a natural language processing result based on a natural language query; and a schema encoder 242 model that outputs a schema relationship.

The schema encoder 242 model may output a schema semantic expression as a schema relationship based on an input schema graph.

The language model 244 , which is a natural language processing model, may receive a preprocessing result of a natural language query and output semantic expressions of natural language elements (eg, parsed words, morphemes, and tokens) as a natural language processing result.

A natural language query processing apparatus according to an embodiment of the present invention includes a first cross-attention layer 246 generating a first cross-attention result indicating a degree of relevance of a result of natural language processing with a schema relationship; and a second cross-attention layer 248 generating a second cross-attention result indicating a degree of relevance of the schema relationship with the natural language processing result.

The language model 244, which is a natural language processing model, may include a first inner attention layer having a self-attention function therein, and the schema encoder 242 model has at least one of self-attention and cross-attention functions therein. A second inner attention layer may be included.

5 is a conceptual diagram illustrating an internal structure of a natural language query processing apparatus according to an embodiment of the present invention, when updating of internal parameters of a language model is blocked.

The natural language query processing apparatus according to an embodiment of the present invention may block updating of internal parameters of the natural language processing model/language model 244 based on the output of the first internal attention layer.

The natural language query processing apparatus according to an embodiment of the present invention may block updating of internal parameters of the natural language processing model/language model 244 based on a cross-attention result.

Referring to FIG. 5 , since the natural language processing process and the schema encoding process are separated, only the schema encoder 242 model can be pre-learned independently of the natural language processing process. This is a big difference between the present invention and the prior art, and is also a factor that can further advance the natural language query processing technique according to the embodiment of the present invention.

By separating the pre-learning process of the schema encoder 242 model from the natural language processing process, the schema encoder 242 model can be unsupervised by generating a schema from information in a lower database (table) that can be easily obtained from the web. By pre-learning the schema encoder 242 model, a problem in the prior art that the schema input form does not match the input form of the pre-learned model can be solved.

Two learning processes for pre-learning may use masked column modeling and/or table relationship prediction.

Similar to mask language modeling, mask column modeling capable of learning associations in tables can be performed by masking table column names with a probability of 15% and then predicting the corresponding column names. At this time, as the semantic expression used to predict the name of the column, the semantic expression obtained through the fully-connected layer of the second layer after passing through BERT can be used, and the name of the column masked by the model It can be controlled to solve a classification problem that predicts .

The table relationship prediction technique, which can be performed as a learning to understand the connectivity between tables, masks the relationship information entering the relationship recognition transformer with a probability of 15%, and models the classification problem to match which relationship should be placed in the corresponding position. This can be performed by being controlled to solve.

A more specific process of masked column modeling and/or table relationship prediction will be described with reference to FIGS. 10 to 13 to be described later.

Referring back to FIGS. 4 and 5 , the natural language query processing apparatus according to an embodiment of the present invention may generate an encoding result by individually considering the natural language relationship and the schema relationship of the natural language query in the encoding step. The language model 244 and the schema encoder 242 may independently calculate semantic representations for their respective inputs.

In order to effectively create a structured query based on a natural language query, it is important to accurately recognize the relationship between a schema and a natural language. To this end, the natural language query processing apparatus according to an embodiment of the present invention may use cross-attention.

A natural language query processing apparatus according to an embodiment of the present invention may be configured so that information between two inputs (a result of natural language processing and a schema relationship) respectively encoded using cross attention can be referenced and utilized. An equation for calculating an attention score in cross attention may include r denoting relationship information. In this case, a relationship calculated through schema linking may be used as relationship information r. The final semantic expression of the natural language processing relationship and the schema relationship can be calculated using cross attention.

Through this configuration, the embodiment of the present invention can more accurately calculate the natural language processing result and the semantic expression for the database schema necessary for translating a natural language query into a standardized/structured query such as an SQL query, thereby improving the performance of the translation system. there is.

6 is a conceptual diagram illustrating an embodiment further including a decoder layer generating a structured query based on a result of cross attention as an internal structure of a natural language query processing apparatus according to an embodiment of the present invention.

The apparatus for processing a natural language query according to an embodiment of the present invention may further include a decoder layer 250 that generates a structured query based on a cross attention result.

Referring to the embodiments of FIGS. 4 to 6 , in order to generate a structured query, a semantic expression may be additionally encoded by cross-referencing a natural language processing result in a natural language query and a schema relationship that may not be directly included in the natural language query. . The schema relationship may include both information included in the natural language query and information not included in the natural language query.

Sub-database (table/column) information to be focused on for natural language query translation among schema relationships may be obtained by using the second intersection attention 248 of the natural language processing result for the schema relationship. As a result, some tables/columns from the entire schema can be selected and used meaningfully.

Information on which of the natural language elements (parsing result, word, morpheme, token) constituting the natural language query is related to which table/column by using the first cross attention 246 of the schema relationship for the natural language processing result can be obtained. At this time, a structured query may be created using a semantic correspondence between words in a natural language query and tables/columns in a database.

7 is an operational flowchart illustrating a natural language query processing method executed by a natural language query processing apparatus according to an embodiment of the present invention.

A natural language query processing method according to an embodiment of the present invention may be executed by a computing system including a processor. A natural language query processing method according to an embodiment of the present invention includes receiving a natural language query input by a user 100 (S310); and generating a structured query based on the natural language query (S320), wherein generating the structured query (S320) includes obtaining a natural language processing result for the natural language query (S340); Obtaining a schema relationship extracted based on a relationship between lower databases (tables/columns) in the database related to the natural language query (S330)); Obtaining a cross-attention result generated between the natural language processing result and the schema relationship (S350); and generating a structured query using the cross attention result (S360).

FIG. 8 is an operation flow chart showing steps S350 of FIG. 7 and the processes before and after it in more detail.

Obtaining a cross-attention result (S350) may include obtaining a first cross-attention result indicating a degree of relevance of a result of natural language processing with a schema relationship (S352); and acquiring a second cross-attention result indicating the degree of relevance of the schema relationship with the natural language processing result (S354).

In the step of obtaining a schema relationship ( S330 ), a schema semantic expression may be obtained as a schema relationship based on an input schema graph. In step S340 of obtaining a natural language processing result, a preprocessing result of a natural language query is input, and semantic expressions of natural language elements (parsing results, words, morphemes, and tokens) may be obtained as the natural language processing result.

Obtaining the natural language processing result (S340) may be performed in a state in which updating of parameters within the natural language processing model based on the output of the internal attention layer is blocked.

Obtaining the natural language processing result ( S340 ) may be performed in a state in which updating of internal parameters of the natural language processing model based on the cross attention result is blocked.

9 is an operational flowchart illustrating a natural language query processing method according to another embodiment of the present invention.

A natural language query processing method executed by a computing system including a processor according to an embodiment of the present invention provides a one-to-one correspondence between sub-databases in a database related to a natural language query, and Extracting as one of one-to-many correspondence (S430); and generating a structured query from the natural language query based on the schema relationship (S460). The natural language query processing method according to an embodiment of the present invention may further include receiving a natural language query (S410).

The natural language query processing method according to an embodiment of the present invention may further include predicting a schema relationship as one of a mandatory relationship and an optional relationship through comparison between lower databases.

A natural language query processing method according to an embodiment of the present invention includes encoding a schema relationship using a first natural language model corresponding to a first sub-database and a second natural language model corresponding to a second sub-database among sub-databases. can include more.

The natural language query processing method according to an embodiment of the present invention may further include pre-encoding at least a portion of a schema relationship by using mask column modeling for a portion of lower databases.

The natural language query processing method according to an embodiment of the present invention may further include obtaining a natural language processing result based on the natural language query. In this case, the generating of the structured query may include obtaining a result of natural language processing and a cross-attention result generated between schema relationships; and generating a structured query using the cross attention result.

The natural language query processing method according to an embodiment of the present invention may further include obtaining a natural language processing result based on a natural language query using a pre-learned language model. In this case, the step of obtaining the natural language processing result may be performed in a state in which updating of internal parameters of the language model is blocked.

The natural language query processing method according to the embodiment of FIG. 9 may include encoding a transformer neural network-based natural language query and a relational database into a meaningful expression.

At this time, the method may use the same encoder as in the prior art, but as a result, a method of pre-learning by newly creating natural language query and SQL pair data may be proposed.

Problems revealed in the prior art include lack of training data for natural language queries and schema pairs, limitations on the number of input tokens of natural language processing models (not applicable to schemas with a large number of columns and tables), and inefficient relationship information between inputs. In order to solve the problem of use, in an embodiment of the present invention, a process of encoding a natural language query and a schema may be separated. Since a natural language query is similar to a general natural language sentence, a pretrained language model can be used without additional training, and a semantic expression of a natural language element can be calculated.

In order to analyze the schema, a semantic expression for each table and columns belonging to the table is first calculated, and a final semantic expression may be calculated by interacting with other table information using a relational recognition transformer neural network. Detailed relationship information between the tables used at this time can be predicted in the pre-processing process.

In the prior art, since a natural language query and a schema are simultaneously encoded as one model, there is a problem in that training data of a natural language query and schema pair is required to pre-learn the corresponding model. However, since the proposed model separates the natural language query and the schema encoding process/module, the schema encoding module can be pre-learned independently.

In an embodiment of the present invention, schema information may be generated from a table extracted from generalized data such as the Internet. Schema information may be unsupervised learning through mask language modeling. By independently pre-learning only the schema encoder 242 module, the problem of mismatch between the pre-learned model and the input form, which is a problem in the prior art, can be solved.

In the embodiment of the present invention, since the language model 244 encodes a natural language query and table information, a schema composed of a plurality of tables can be processed without exceeding the length limit.

In the embodiment of the present invention, the problem of interaction without considering the connection relationship between tables, which is a problem revealed in the prior art, does not occur because the schema is encoded using a relational recognition transformer that encodes in table units first and considers the association of each table. .

According to an embodiment of the present invention, when a natural language query and a schema are given, relational information between input values may be extracted through a preprocessing process.

The pre-processing process may include schema linking and schema relation extraction.

In schema linking, relational information between words w in a natural language query and table t or column c can be obtained.

In schema relation extraction, relations between tables found through the inclusion relation between tables and columns, primary-foreign keys in the schema graph, and table records can be extracted.

After the preprocessing is finished, the natural language query is encoded by the pretrained language model 244, and a natural language processing result may be generated. A schema relationship may be obtained by the pretrained schema encoder 242 . A natural language processing result may be calculated as a semantic expression of a natural language element, and a schema relation may be calculated as a semantic expression of a schema.

Table relationships extracted in the preprocessing step can be used when calculating the schema semantic expression.

When the calculation of the semantic expression for each input is primarily completed, the final semantic expression of the two inputs may be calculated through information exchange between the natural language query and the schema through the relation recognition transformer. At this time, information on which word of the natural language query is related to which schema entity may be obtained based on information extracted in schema linking, which is a preprocessing step.

At this time, the schema linking process, which is a part of the preprocessing process, is a well-known technique, KR 10-2277787, "Method of Predicting Columns and Tables Used in Translating SQL Queries from Neural Network-Based Natural Language" and KR 10-2345568, "Natural Language Words in Database Columns and It can be executed using a 'configuration for matching by comparing a natural language sentence and a database word by word', which is at least a part of the contents disclosed in "How to connect to a table".

10 is a conceptual diagram illustrating an embodiment of a schema linking process when a natural language query and a schema related to automobiles are given in a natural language query processing method performed by a natural language query processing apparatus according to an embodiment of the present invention.

10 illustrates an example of schema linking when a natural language query and a schema related to automobiles are given.

Referring to FIG. 10 , a schema linking step may be performed as an input text preprocessing process in a natural language processing apparatus and/or an apparatus for encoding a natural language query and a relational database based on a transformer neural network according to an embodiment of the present invention into a meaningful expression.

Schema linking is a process of solving a problem quite similar to entity linking, in which an entity to be linked is found in a schema rather than a knowledge graph. That is, a word in a natural language query is linked to a table or column that the word refers to. As shown in FIG. 10, words of a natural language query may be connected to one or more columns/tables. Schema linking is a scheme that has already been used in various studies, and in the present invention, a prior art schema linking algorithm can be used at least partially as needed.

In schema linking, string comparison is performed with each table t in the schema for each word w of the given natural language query L. At this time, if completely matching words such as cylinders and horsepower in FIG. 10 are found, they are connected. However, if there is no complete match, it finds a partial match and connects all found objects. At this time, it can be connected based on words rather than partial matching strings. If there is no partial matching table, the same operation is repeated for the column.

11 illustrates that in a natural language query processing method performed by a natural language query processing apparatus according to an embodiment of the present invention, one natural language query can be translated into different structured queries according to relationships between sub-databases (tables) in a database. It is a conceptual diagram to explain.

Referring to FIG. 11 , table relationship extraction is a process of finding a relationship between two tables and is a necessary process for accurate query translation. As shown in FIG. 11 , for the same natural language query, SQL to be translated may vary according to the relationship between tables in the schema. In the case of SQL1, the translation is correct when the relationship between the student and has_pet tables is one-to-one, and SQL2 is the correct translation when the relationship is one-to-many.

That is, even for the same natural language query, a structured query when the relationship between student and pet (has_pet) is one-to-one and a structured query when the relationship between student and pet is one-to-many are not the same. Therefore, in order to generate a correct structured query, more schema relationships that are not directly included in the natural language query are required.

12 is a conceptual diagram illustrating an embodiment of extracting relationships between lower databases (tables) by utilizing database record values in a natural language query processing method performed by a natural language query processing apparatus according to an embodiment of the present invention.

Referring to FIG. 12 , in one embodiment of the present invention, a database record value is used as a method of extracting a relationship between two tables.

When there is a column C1 set as the primary key and column C2 set as the foreign key, the values corresponding to each column in the table record are called V1 and V2. In this case, if there is an overlapping value in V2, a one-to-many relationship may be extracted, and if there is no overlapping value, a one-to-one relationship may be extracted.

In addition, a mandatory relationship can be extracted/predicted if the set of V1 and the set of V2 match, and an optional relationship can be extracted/predicted if they do not match.

Referring to FIG. 12, it can be seen that the relationship between id of the student table and student_id of the has_pet table is one-to-many optional. The relationship between the has_pet table and the pet table is one-to-one mandatory.

13 is a conceptual diagram illustrating an example of schema encoding in a natural language query processing method performed by a natural language query processing apparatus according to an embodiment of the present invention.

Referring to FIG. 13 , in an embodiment of the present invention, a natural language query encoding step and a schema encoding step are performed as a semantic expression calculation process.

For convenience of description, it is assumed that the encoding of natural language queries proceeds through a pretrained language model BERT.

Natural language sentences are separated into tokens through byte-pair tokenization. At this time, the vocabulary of the token uses the same as BERT. Then, the semantic expression for each token can be obtained by inputting the natural language query converted into tokens as input to BERT. At this time, a semantic expression for each word w in the sentence L may be calculated. When a word is divided into several tokens, a semantic expression for each word may be calculated by averaging the semantic expressions of each token.

Schema encoding is divided into two steps as shown in FIG. 13 . First, using the pre-trained language model BERT, the semantic representation for each table t in S in the schema can be computed independently. In order to put table information into BERT, a process of flattening table information into sentences may be performed. As table information, the name of the table and the names of columns belonging to the table can be used. The process of unfolding table information into sentences is as follows.

First, you can connect by tokenizing the table name after a special token called [CLS] (classifier). After that, all column names belonging to the table can be tokenized and concatenated after the table name. At this time, columns are separated using a special token called [SEP] (separator) between columns. Connected tokens are input into BERT to calculate the semantic expression for each table and column, and based on the relationship information between tables and between tables/columns extracted in the preprocessing step, the schema is encoded again using a relationship-aware transformer. can

An embodiment of the present invention may be a computer program stored in a computer-readable storage medium for implementing a method of encoding a transformer neural network-based natural language query and a relational database into a meaningful expression, and/or a method of processing a natural language query.

An embodiment of the present invention may be a computer readable recording medium for implementing a program of a method of encoding a transformer neural network-based natural language query and a relational database into a meaningful expression and/or a method of processing a natural language query.

According to the embodiments of FIGS. 1 to 13 , a correct structured query may be generated by extracting relationships between tables.

According to the embodiments of FIGS. 1 to 13 , the size of input data may not be limited by separating learning of natural language processing and learning of schema encoding. Therefore, it is possible to implement an automatic translation function for natural language queries for databases using hundreds of columns that companies actually need.

According to the embodiments of FIGS. 1 to 13 , natural language encoding using a pretrained language model and schema encoding using a relation recognition transformer may be proposed.

According to the embodiments of FIGS. 1 to 13 , since language models can also be individually applied for each table, encoding can be performed without limitation on the number of tables.

According to the embodiments of FIGS. 1 to 13 , encoding for generating a structured query may be performed using relationship information between predicted tables.

According to the embodiments of FIGS. 1 to 13 , information on a lower level database to be focused on in relation to a natural language in a schema may be obtained using cross attention.

According to the embodiments of FIGS. 1 to 13 , cross attention can be used to obtain information on which words in natural language to focus on (information on which words are highly related to which schema).

According to the embodiments of FIGS. 1 to 13 , elements to be focused on in each of the natural language and schema can be recognized using cross attention, so that the process of creating a structured query can be advanced.

14 is a block diagram illustrating a generalized natural language processing device or computing system capable of performing at least a part of the processes of FIGS. 1 to 13 according to an embodiment of the present invention.

As described above, in the embodiments of FIGS. 1 to 3, the dialogue manager 220, the query translator 240, and the response engine 260 are electronically connected to the processor and memory and can be controlled or managed by the processor. .

Although omitted in the drawings in the embodiments of FIGS. 4 to 13 , a processor and a memory may be electronically connected to each component, and operations of each component may be controlled or managed by the processor.

At least some processes of the natural language processing method according to an embodiment of the present invention may be executed by the computing system 1000 of FIG. 14 .

Referring to FIG. 14 , a computing system 1000 according to an embodiment of the present invention includes a processor 1100, a memory 1200, a communication interface 1300, a storage device 1400, an input interface 1500, and an output It may be configured to include an interface 1600 and a bus 1700.

The computing system 1000 according to an embodiment of the present invention includes at least one processor 1100 and instructions instructing the at least one processor 1100 to perform at least one step. It may include a memory (memory) 1200 for storing. At least some steps of the method according to an embodiment of the present invention may be performed by the at least one processor 1100 loading instructions from the memory 1200 and executing them.

The processor 1100 may mean a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor on which methods according to embodiments of the present invention are performed.

Each of the memory 1200 and the storage device 1400 may include at least one of a volatile storage medium and a non-volatile storage medium. For example, the memory 1200 may include at least one of a read only memory (ROM) and a random access memory (RAM).

Also, the computing system 1000 may include a communication interface 1300 that performs communication through a wireless network.

In addition, the computing system 1000 may further include a storage device 1400, an input interface 1500, an output interface 1600, and the like.

In addition, each component included in the computing system 1000 may be connected by a bus 1700 to communicate with each other.

For example, the computing system 1000 of the present invention includes a communicable desktop computer, a laptop computer, a notebook, a smart phone, a tablet PC, and a mobile phone. (mobile phone), smart watch, smart glass, e-book reader, PMP (portable multimedia player), portable game device, navigation device, digital camera, DMB (digital It may be a multimedia broadcasting) player, digital audio recorder, digital audio player, digital video recorder, digital video player, personal digital assistant (PDA), and the like. .

The operation of the method according to the embodiment of the present invention can be implemented as a computer readable program or code on a computer readable recording medium. A computer-readable recording medium includes all types of recording devices in which information readable by a computer system is stored. In addition, computer-readable recording media may be distributed to computer systems connected through a network to store and execute computer-readable programs or codes in a distributed manner.

In addition, the computer-readable recording medium may include hardware devices specially configured to store and execute program instructions, such as ROM, RAM, and flash memory. The program command may include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine code generated by a compiler.

Although some aspects of the present invention have been described in the context of an apparatus, it may also represent a description according to a corresponding method, where a block or apparatus corresponds to a method step or feature of a method step. Similarly, aspects described in the context of a method may also be represented by a corresponding block or item or a corresponding feature of a device. Some or all of the method steps may be performed by (or using) a hardware device such as, for example, a microprocessor, programmable computer, or electronic circuitry. In some embodiments, at least one or more of the most important method steps may be performed by such a device.

In embodiments, a programmable logic device (eg, a field programmable gate array) may be used to perform some or all of the functions of the methods described herein. In embodiments, a field-programmable gate array may operate in conjunction with a microprocessor to perform one of the methods described herein. Generally, methods are preferably performed by some hardware device.

Although the above has been described with reference to preferred embodiments of the present invention, those skilled in the art can variously modify and change the present invention without departing from the spirit and scope of the present invention described in the claims below. You will understand that you can.

Claims (20)

processor;
including,
The processor
Receive a natural language query entered by a user;
Create a structured query based on the natural language query;
When the processor generates the structured query based on the natural language query,
a natural language processing result for the natural language query; and
a schema relationship extracted based on a relationship between sub-databases within a database related to the natural language query;
Natural language query processing apparatus for generating the structured query by using a mutually generated cross-attention result of According to claim 1,
The processor queries the structured query using a first cross-attention result indicating the degree of relevance of the natural language processing result with the schema relationship, and a second cross-attention result indicating the degree of relevance of the schema relationship with the natural language processing result. create,
The natural language query processing apparatus of claim 1 , wherein the cross-attention result includes the first cross-attention result and the second cross-attention result. According to claim 1,
a natural language processing model generating the natural language processing result based on the natural language query; and
a schema encoder model outputting the schema relationship;
Further comprising a natural language query processing device. According to claim 3,
The schema encoder model outputs a schema semantic expression as the schema relationship based on an input schema graph,
wherein the natural language processing model receives a preprocessing result of the natural language query and outputs a semantic expression of a natural language element as the natural language processing result. According to claim 3,
a first cross-attention layer that generates a first cross-attention result indicating a degree of relevance of the natural language processing result with the schema relationship; and
a second cross-attention layer generating a second cross-attention result indicating a degree of relevance of the schema relationship with the natural language processing result;
Including more,
The natural language query processing apparatus of claim 1 , wherein the cross-attention result includes the first cross-attention result and the second cross-attention result. According to claim 3,
The natural language processing model includes a first inner attention layer having a self-attention function therein,
The natural language query processing apparatus of claim 1, wherein the schema encoder model includes a second inner attention layer having at least one function of self-attention and cross-attention. According to claim 6,
The processor
The natural language query processing apparatus for blocking updating of parameters within the natural language processing model based on an output of the first inner attention layer. According to claim 3,
The processor
A natural language query processing device that blocks updating of parameters within the natural language processing model based on a result of the cross attention. According to claim 3,
a decoder layer generating the structured query based on the cross attention result;
Further comprising a natural language query processing device. In the natural language query processing method executed by a computing system including a processor,
Receiving a natural language query input by a user; and
generating a structured query based on the natural language query;
including,
Generating the structured query,
obtaining a natural language processing result for the natural language query;
obtaining a schema relationship extracted based on a relationship between sub-databases in a database related to the natural language query;
obtaining a cross-attention result generated between the natural language processing result and the schema relationship; and
generating the structured query using the cross attention result;
Including, natural language query processing method. According to claim 10,
Obtaining the cross attention result,
obtaining a first cross-attention result indicating a degree of relevance of the natural language processing result with the schema relationship; and
obtaining a second cross-attention result indicating a degree of relevance of the schema relationship with the natural language processing result;
Including, natural language query processing method. According to claim 10,
Obtaining the schema relationship,
Obtaining a schema semantic expression as the schema relationship based on an input schema graph;
Obtaining the natural language processing result,
The natural language query processing method of receiving a preprocessing result of the natural language query and obtaining a semantic expression of a natural language element as the natural language processing result. According to claim 10,
Obtaining the natural language processing result,
A natural language query processing method performed in a state in which updating of parameters in a natural language processing model based on an output of an internal attention layer is blocked. According to claim 10,
Obtaining the natural language processing result,
The natural language query processing method, which is performed in a state in which updating of parameters in the natural language processing model based on the result of the cross attention is blocked. In the natural language query processing method executed by a computing system including a processor,
extracting a schema relationship between sub-databases in a database related to a natural language query as one of one-to-one correspondence and one-to-many correspondence; and
generating a structured query from the natural language query based on the schema relationship;
Including, natural language query processing method. According to claim 15,
predicting the schema relationship as one of a mandatory relationship and an optional relationship through comparison between the lower databases;
Natural language query processing method further comprising a. According to claim 15,
encoding the schema relationship using a first natural language model corresponding to a first sub-database and a second natural language model corresponding to a second sub-database among the sub-databases;
Natural language query processing method further comprising a. According to claim 15,
pre-encoding at least a portion of the schema relationship by using mask column modeling for a portion of the sub-database;
Natural language query processing method further comprising a. According to claim 15,
obtaining a natural language processing result based on the natural language query;
Including more,
Generating the structured query,
obtaining a cross-attention result generated between the natural language processing result and the schema relationship; and
generating the structured query using the cross attention result;
Including, natural language query processing method. According to claim 15,
obtaining a natural language processing result based on the natural language query by using a pre-learned language model;
Including more,
The natural language query processing method of claim 1 , wherein the obtaining of the natural language processing result is performed in a state in which updating of internal parameters of the language model is blocked.

Images