Conversational Tabular Data Access Using Alias-Based Row Text
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conversational AI systems face challenges in processing and generating coherent natural language answers from large volumes of tabular data due to the structured nature of tables, leading to computational strain and inaccuracies in mapping natural language queries to SQL operations or hard-coded phrases, which fail to capture nuanced relationships in the data.
Innovation Solution
A method involving a data server that generates a dummy table and alias table using a large language model to create a dummy row text representation, replacing alias row data with actual row data to generate a row text representation, enabling conversational access to tabular data without sharing the actual data with the model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the entire tabular dataset is input into a model to generate natural language answers, then the accuracy of answers improves, but computational resources are strained and processing speed decreases
Solution Approach 1:
The patent segments the tabular data processing into multiple stages: (1) generating dummy data that preserves structural patterns, (2) using this dummy data to create text representations, and (3) applying these representations to answer queries. This segmentation allows the model to work with compressed representations rather than full datasets, improving speed while maintaining answer accuracy through pattern preservation.
Solution Approach 2:
The patent creates dummy data copies that replicate the structural patterns and relationships of the original tabular data without containing actual sensitive information. These dummy copies serve as proxies for the full dataset, enabling the model to learn patterns and generate accurate answers while significantly reducing computational resource requirements and processing time.
2Ease of manufacture
If SQL operations are used to map natural language queries to tabular data, then the technical implementation is straightforward, but the accuracy and reliability of answers decrease
Solution Approach 1:
The patent introduces dummy data and text representations as intermediary elements between natural language queries and the tabular database. Instead of directly mapping queries to SQL operations, the system uses dummy-generated text representations that capture semantic relationships, serving as a mediator that improves answer accuracy while maintaining implementation feasibility through automated generation processes.
3Device complexity
If hard-coded phrases with concatenated column names and row values are generated, then the implementation complexity is reduced, but the semantic association and accuracy of data representation deteriorate
Solution Approach 1:
The patent performs preliminary actions by generating dummy data and creating text representations before actual query processing. This preliminary generation of structured text representations with proper semantic associations establishes a foundation for accurate data representation, avoiding the need for complex real-time processing while maintaining high semantic accuracy in the final outputs.
Data Source
AI summary
Methods, non-transitory computer readable media, and a data server that assist with enabling conversational access to tabular data includes determining in response to a user input, tabular data comprising a header row with header data in each column and one or more table data rows with row data in each column. A first prompt is provided to a large language model to generate a dummy table comprising the header row and a dummy row with dummy row data in each column and a dummy table is received. An alias table comprising the header row and an alias row with alias row data in each column is generated. A second prompt is provided to the large language model to generate a dummy row text representation of the dummy row data, wherein the dummy row text representation includes the alias row data inserted as placeholders of the dummy row data and the dummy row text representation is received. A row text representation is generated for each of the one or more table data rows by replacing the alias row data in the dummy row text representation with row data of corresponding ones of the one or more table data rows.