Question-answering method, method and apparatus for training table question-answering model, device, and medium

By converting table documents into coding formats and building training data, the table question and answer model is trained, which solves the problem of difficulty in analyzing and low utilization of table information in natural language models, and realizes more efficient table information utilization and complex table question and answer capabilities.

WO2025149813A1PCT designated stage expired Publication Date: 2025-07-17CLOUD INTELLIGENCE ASSETS HOLDING (SINGAPORE) PTE LTD
View PDF 3 Cites 4 Cited by

Patent Information

Application Number
PCT/IB2024/062731
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2024-01-10
Filing Date
2024-12-17
Publication Date
2025-07-17

AI Technical Summary

Technical Problem

When processing table information, existing natural language models have problems such as difficulty in analyzing and low information utilization, especially the lack of complex structure table and multi-table Q&A capabilities.

Method used

By converting the table document into encoding format (such as HTML or LaTeX), the table content and structure information are extracted, and training data is constructed based on the preset question-and-answer guide template, the table question-and-answer model is trained, and the question-and-answer results corresponding to natural language questions are generated.

Benefits of technology

Improve the utilization and recall rate of table information in the question-and-answer model, improve the ability of complex tables and multi-table question-and-answer questions and answers, and ensure the integrity of the table content structure and logical relationship.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure IB2024062731_17072025_PF_FP_ABST
    Figure IB2024062731_17072025_PF_FP_ABST
Patent Text Reader

Abstract

The present disclosure relates to the technical field of artificial intelligence, and provides a question-answering method, a method and an apparatus for training a table question-answering model, a device, and a medium. The method comprises: selecting, from a plurality of candidate tables, a target table that contains table information matching a natural language question, wherein the candidate tables are in encoded format; on the basis of the natural language question and the table information of the target table, generating a corresponding table question; and inputting the generated table question into a pre-trained table question-answering model to obtain a question-answering result corresponding to the natural language question. The technical solution of the present disclosure can improve the utilization of table information in table question-answering models and enhance the capability for question answering on complex tables and multiple tables.
Need to check novelty before this filing date? Find Prior Art

Claims

Claims 1. A table-based question answering method, comprising: Select a target table whose table information matches the natural language question from multiple candidate tables, where the candidate tables are in encoded format; generate a corresponding table question based on the natural language question and the table information of the target table; input the generated table question into a pre-trained table question-answering model to obtain an answer result corresponding to the natural language question.

2. The method according to claim 1 further comprises: Parse the table in encoded format from the table document to obtain the candidate tables; Extract the table information from the candidate tables, where the table information includes table content information and table structure information.

3. The method according to claim 2, wherein Parse the table in encoded format from the table document to obtain the candidate tables, including: parsing the table in encoded format from the table document; performing formatting processing on the encoded format table to obtain the candidate tables, where the formatting processing includes removing table styles and / or removing preset tags, and the preset tags are used to define at least one of page layout, page style, and interaction method.

4. The method according to claim 2, wherein The table structure information includes codes for characterizing the table structure; the table content information includes at least one of a table title, a table introduction, cell fields, and cell content.

5. The method according to claim 2, wherein The table content information includes a table title and / or a table introduction. Extracting the table content information from the candidate tables includes: determining the associated node corresponding to the candidate table in the table document, where the associated node includes at least one of a sibling node, a previous node, and a next node; searching for the table title and / or the table introduction from the associated nodes of the candidate table.

6. The method according to any one of claims 1 to 5, wherein Selecting a target table whose table information matches the natural language question from multiple candidate tables in encoded format, including: using multiple matching algorithms to select candidate tables whose table information matches the natural language question to obtain multiple candidate tables to be recalled; sorting the multiple candidate tables to be recalled according to the matching degree between the table information and the natural language question; selecting the target table from the multiple candidate tables to be recalled based on the sorting result.

7. The method according to any one of claims 1 to 6, further comprising: Construct the training data of the table question-answering model based on a preset question-answering guidance template, where the question-answering guidance template includes a table question template and an answer result template, and the training data includes table questions that conform to the table question template and answer results that conform to the answer result template; use the training data to train the table question-answering model.

8. The method according to claim 7, wherein, Construct the training data of the table question-answering model, including: using a pre-trained natural language processing model to generate table questions for the candidate tables that conform to the table question template; inputting the table questions for the candidate tables that conform to the table question template into the natural language processing model to generate answer results that conform to the answer result template.

9. The method according to claim 7 or 8, wherein Generate a corresponding table question based on the natural language question and the table information of the target table, including: generate a table question corresponding to the natural language question and the target table and conforming to the table question template based on the natural language question and the table information of the target table.

10. A training method for a table question-answering model, comprising: Construct training data for the table question-answering model based on a preset question-answering guidance template, where the question-answering guidance template includes a table question template and a question-answering result template, and the training data includes table questions conforming to the table question template and question-answering results conforming to the question-answering result template; use the training data to train the table question-answering model, where the table question-answering model is used to obtain a question-answering result corresponding to the natural language question according to the input table question, and the table question is generated based on the natural language question and a table matching the natural language question.

11. A table-based question answering device, comprising: A target table matching module, configured to select a target table whose table information matches the natural language question from multiple candidate tables, where the candidate tables are in an encoded format; a table question generation module, configured to generate a corresponding table question based on the natural language question and the table information of the target table; a question-answering result generation module, configured to input the generated table question into a pre-trained table question-answering model to obtain a question-answering result corresponding to the natural language question.

12. A training device for a tabular question-answering model, comprising: A training data construction module, configured to construct training data for the table question-answering model based on a preset question-answering guidance template, where the question-answering guidance template includes a table question template and a question-answering result template, and the training data includes table questions conforming to the table question template and question-answering results conforming to the question-answering result template; a training module, configured to use the training data to train the table question-answering model, where the table question-answering model is used to obtain a question-answering result corresponding to the natural language question according to the input table question, and the table question is generated based on the natural language question and a table matching the natural language question.

13. An electronic device, including a memory, a processor, and a computer program stored on the memory, where the processor implements the method according to any one of claims 1 to 10 when executing the computer program.

14. A computer-readable storage medium, in which a computer program is stored, and the 19 computer program, when executed by a processor, implements the method according to any one of claims 1 to 10.

15. A computer program product, including a computer program, where the computer program, when executed by a processor, implements the method according to any one of claims 1 to 10.

Citation Information

Patent Citations

  • Mode link processing method and device and computer readable storage medium

    CN114661963A

  • Data query method and device, electronic equipment and storage medium

    CN116303537A

  • Table question and answer processing method, device and equipment

    CN117332054A

Cited By

  • Intelligent question and answer enhancement method and system based on RAG technology

    CN120723877A

  • Intelligent question answering method and system based on large model

    CN121029955A

  • Multi-modal problem solving method and multi-modal problem solving system

    CN121328733A

  • Table question and answer task capability enhancement processing method of large language model

    CN121598963A