Multi-Score Answer Mining for High Recall Retrieval

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing general answer mining technologies have low recall, failing to accurately provide correct texts as answers due to pre-screening, leading to incorrect selections.

Innovation Solution

A method and apparatus that acquire questions, determine candidate articles and texts, and assign multiple scores to evaluate matching degrees, using a machine learning model to comprehensively select answers based on a score set, ensuring accurate and correct answer retrieval.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If pre-screening is applied to filter candidate articles and texts, then the efficiency of answer selection is improved, but the recall deteriorates causing correct answers to be incorrectly filtered

Engineering Contradiction:
Improveefficiency of answer selectionVSAvoidrecall of correct answers
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent changes the evaluation parameter from a single binary filter (pass/fail) to a multi-dimensional scoring system. Each candidate text receives multiple scores from different models (semantic matching score, keyword matching score, position score, etc.), and the final selection is based on comprehensive scoring rather than simple threshold filtering. This resolves the contradiction by maintaining high recall while improving efficiency through structured evaluation.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent combines multiple evaluation models and scoring mechanisms into a composite answer selection system. Different models (semantic matching, keyword matching, position-based) are combined to form a comprehensive evaluation framework. This composite approach ensures that correct answers are not filtered out by any single model's limitations while maintaining efficient selection through the integrated scoring system.

Inventive Principle:
Principle #40Composite materials

2Measurement precision

If multiple scoring models are used to evaluate candidate texts, then the correctness of answer selection is improved, but the system complexity increases

Engineering Contradiction:
Improvecorrectness of answer selectionVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the answer selection task into multiple independent scoring components, each handled by a separate model (semantic matching model, keyword matching model, position scoring model). Each model focuses on a specific aspect of answer quality, and their results are aggregated. This segmentation improves measurement precision by specializing each model while managing complexity through modular design.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent creates a universal scoring framework that can accommodate multiple different evaluation models. The framework provides a standardized interface for integrating various scoring mechanisms (semantic, keyword, position, etc.) and aggregating their results. This multi-functional design improves correctness by allowing diverse evaluation approaches while controlling complexity through a unified structure.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentUS11366819B2Method, apparatus, electronic device and computer readable medium for obtaining answer to question
Publication Date: 2022.06.21 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • US11366819B2 patent drawing
  • US11366819B2 patent drawing
  • US11366819B2 patent drawing

AI summary

A method for obtaining an answer to a question is provided. The method may include: acquiring a question; determining at least a part of articles in a preset article database as candidate articles, and determining first scores of the candidate articles respectively, the first score of any of the candidate articles representing a matching degree between the candidate article and the question; determining at least a part of texts in each of the candidate articles as candidate texts, and determining second scores of the candidate texts respectively, the second score of any of the candidate texts representing a matching degree between the candidate text and the question; and determining at least a part of the candidate texts as the answer based on a score set of each of the candidate texts, the score set of any of the candidate texts including the second score and the first score.