Table Query Answering Using Cell Relevance Scores
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine-learning models, particularly large language models, struggle with accurately answering questions from tables due to their susceptibility to noise and inability to process the underlying structure of tables, leading to inaccurate results and inefficient use of computational resources.
Innovation Solution
A system that calculates relevance scores for table cells based on a query, using a combination of unsupervised and weakly-supervised methods to focus machine-learning models on relevant content, suppressing irrelevant information and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine-learning models process entire table content, then comprehensive search results are generated, but accuracy decreases due to noise and computational resources are wasted
Solution Approach 1:
The patent segments the table content into individual cells and calculates relevance scores for each cell based on the query. This segmentation allows the system to identify and focus on relevant cells while excluding irrelevant ones, thereby improving search result accuracy and reducing computational resource consumption by processing only necessary portions of the table.
Solution Approach 2:
The patent applies local quality by assigning different relevance scores to different cells based on their specific relationship to the query. Cells with higher relevance scores are given more weight in the search result generation, while cells with lower scores are suppressed. This differential treatment improves accuracy by emphasizing relevant information and reduces computational waste by de-emphasizing irrelevant information.
2Reliability
If conventional machine-learning models process entire table content, then all available information is considered, but computational efficiency decreases
Solution Approach 1:
The patent performs preliminary action by calculating relevance scores for table cells before the main search result generation process. This pre-processing step identifies and prioritizes relevant cells in advance, allowing the machine-learning model to focus computational resources on high-value information during the actual search, thereby improving both reliability and processing efficiency.
Solution Approach 2:
The patent applies partial action by processing only the most relevant table cells rather than the entire table content. By using relevance scores to select a subset of cells for detailed processing, the system achieves sufficient search result reliability with reduced computational effort, avoiding the waste associated with processing all table content equally.
3Measurement precision
If relevance scoring is applied to table cells, then focus on relevant content is improved, but system complexity increases
Solution Approach 1:
The patent introduces an intermediary relevance scoring mechanism that sits between the query and the table content processing. This intermediary layer calculates relevance scores based on query-table cell relationships and uses these scores to weight or filter content during search result generation. While this adds a processing step, it systematically improves relevance identification without requiring fundamental architectural changes to the machine-learning model.
Data Source
AI summary
Content relevance based table query answering is described. In one or more examples, a query and a table are received. The table includes a plurality of cells. A plurality of scores for calculated that correspond to the plurality of cells based on the query. One or more machine-learning models are then leveraged to generate a search result from the query, table, and scores, which is presented in a user interface for display.