Supporting Passage Ranking in Question Answering Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In Question Answering (QA) systems, the ranking of supporting passages is hindered by nonlocal features that consider other occurrences of candidate answers, leading to reduced effectiveness in identifying relevant passages.
Innovation Solution
A method that uses only local features to rank supporting passages, where context-dependent scorers assign unique scores to each candidate answer occurrence, and a linear answer scoring model calculates final confidence scores to rank candidate answers and their corresponding passages based on relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If nonlocal features are used to score candidate answers, then the QA system can consider multiple occurrences of candidate answers, but the ability to effectively rank supporting passages is reduced
Solution Approach 1:
The patent segments the scoring process into two independent parts: local feature scoring for passage ranking and nonlocal feature scoring for candidate answer aggregation. By separating these functions, the system can rank passages based solely on local evidence without contamination from other candidate answers, while still allowing nonlocal features to influence final answer selection. This segmentation resolves the contradiction by preventing nonlocal features from interfering with passage ranking effectiveness.
Solution Approach 2:
The patent extracts and removes nonlocal features from the passage ranking process, using only local features for scoring supporting passages. This extraction ensures that passage ranking is based purely on local evidence relevance, eliminating the harmful interference that nonlocal features would cause. The nonlocal features are retained for a separate candidate answer scoring stage, thus solving the contradiction between considering multiple occurrences and maintaining passage ranking precision.
2Measurement precision
If local features are used to rank supporting passages, then passage ranking effectiveness is improved, but the ability to propagate scores across multiple occurrences is lost
Solution Approach 1:
The patent divides the scoring system into two independent modules: a local feature scoring module for passage ranking and a nonlocal feature scoring module for candidate answer aggregation. This segmentation allows local features to effectively rank passages while nonlocal features independently propagate scores across multiple occurrences. The separation eliminates the trade-off by enabling both functions to operate simultaneously without interference.
Solution Approach 2:
The patent introduces an intermediary structure where local features first score passages, then nonlocal features aggregate scores across occurrences to produce final candidate answer scores. This intermediary approach allows score propagation to occur at the candidate answer level rather than interfering with passage-level ranking, thus maintaining both passage ranking effectiveness and score propagation capability.
3Quantity of substance
If multiple occurrences of candidate answers are merged, then redundancy is reduced, but the ability to assign unique local features to each occurrence is lost
Solution Approach 1:
The patent segments the processing pipeline into two stages: first, local features are assigned and scored for each individual occurrence to enable precise passage ranking; second, occurrences are merged and nonlocal features are aggregated to produce final candidate answer scores. This segmentation allows the system to maintain unique local feature assignments for each occurrence while still achieving redundancy reduction through merging at the next stage.
Solution Approach 2:
The patent performs preliminary scoring of local features for each occurrence before merging occurrences. This preliminary action ensures that each occurrence receives its unique local feature assignment and score, enabling accurate passage ranking. Only after this preliminary scoring does the system merge occurrences and apply nonlocal features, thus preserving both local feature accuracy and redundancy reduction.
Data Source
AI summary
Embodiments provide a computer implemented method for ranking supporting passages in a question answering (QA) system. The method includes the following steps: receiving an input question and a plurality of passages, wherein the plurality of passages are extracted from the document corpus by the QA system; identifying a plurality of candidate answer occurrences from the plurality of passages; assigning a plurality of first local features to each candidate answer occurrence; merging the plurality of candidate answer occurrences to generate a group of candidate answers; assigning a plurality of second local features and a plurality of nonlocal features to each candidate answer; ranking the group of candidate answers using a linear answer scoring model; and ranking the plurality of candidate answer occurrences using the linear answer scoring model.