Answer Generation Device Using Token Frequency for Multiple Answers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine reading comprehension models are limited to extracting single answers and struggle to effectively handle multiple answer scenarios, often producing dispersed and non-matching results due to their restrictive nature of outputting only one answer, even when multiple answers are present in the context.
Innovation Solution
The solution involves combining multiple single answer models, each trained for single answer QA, to generate multiple answers by calculating the frequency of token appearance and using threshold values to determine whether a token is part of a single answer, multiple answers, or not an answer at all, thereby allowing for the extraction of single or multiple answers based on their matching rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single answer model is used for machine reading comprehension, then the model structure remains simple and training data requirements are reduced, but the model cannot effectively extract multiple answers and produces dispersed results
Solution Approach 1:
The patent combines multiple single-answer models (first, second, and third models) into an ensemble system. Each model independently processes the input text and question, and their results are aggregated through frequency calculation of answer tokens. This merging approach enables the system to extract multiple answers while maintaining the simplicity of individual single-answer models, resolving the contradiction between adaptability and device complexity.
2Adaptability or versatility
If multiple single answer models are combined to extract multiple answers, then the ability to handle multiple answer scenarios improves, but the processing complexity and computational requirements increase
Solution Approach 1:
The system employs a self-service mechanism where multiple single-answer models independently process the same input without requiring complex coordination or additional training data. Each model serves itself by generating answers based on its own learned parameters, and the aggregation process automatically handles the combination of results through token frequency calculation. This approach improves multiple answer extraction capability while minimizing the increase in processing complexity.
3Ease of manufacture
If conventional single answer models are used, then training data requirements are reduced, but the models produce dispersed and non-matching results when multiple answers are present
Solution Approach 1:
The patent creates multiple copies of single-answer models (first, second, and third models) that are trained on the same or similar training data. These model copies independently process the input and generate answers. By aggregating the results from these model copies through token frequency calculation, the system achieves reliable and consistent multiple answer extraction while maintaining the ease of training individual models on standard datasets.
Data Source
AI summary
An answer generating device includes: an input that receives an input of analysis target data which is data to be questioned and analyzed; a processor; and an output. The processor being configured to, upon input of the question and the analysis target data and by execution of a program, extract answers to the question from the analysis target data using plural single answer models prepared in advance, the answers each being extracted independently for each of the plural single answer models, calculate a frequency of appearance of each of tokens in the extracted answers, and extract a single answer or multiple answers based on the frequency of appearance, and output the single answer or multiple answers to the output.