Non-Factoid Question Answering via Multi-Dimensional Causality Attention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing question-answering systems for non-factoid questions, particularly why-type questions, face challenges in accurately retrieving correct answers due to insufficient consideration of relevance, reasons, and causality in discourse structures, leading to low accuracy.
Innovation Solution
A non-factoid question-answering device that converts questions and answer passages into word semantic vectors, using similarity and attention matrices to weight and combine sentence-level and passage-level representations from multiple contexts, including causality and general viewpoints, to determine the correctness of answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If classifiers trained through machine learning (SVM, neural networks) are used for why-question answering, then the accuracy is improved compared to traditional methods, but the performance is still not sufficiently high
Solution Approach 1:
The patent segments the answer evaluation into three distinct aspects: topic relevance, reason presentation, and causality representation. Each aspect is evaluated separately through specific attention mechanisms (topic attention, reason attention, causality attention), allowing comprehensive assessment of answer quality rather than relying on a single classifier output.
Solution Approach 2:
The patent introduces a new dimensional approach by evaluating answers from multiple perspectives (topic, reason, causality) simultaneously. This multi-dimensional evaluation framework transforms the single-dimension classifier output into a comprehensive multi-faceted assessment, achieving higher performance sufficiency.
2Device complexity
If prior art techniques focus on only part of the answer conditions (topic relevance or structural features), then the system complexity is reduced, but the answer accuracy deteriorates due to insufficient consideration of all three conditions
Solution Approach 1:
The patent divides the complex answer evaluation into three manageable segments: topic relevance evaluation, reason presentation evaluation, and causality representation evaluation. Each segment is handled by a dedicated attention mechanism, making the overall complex system more tractable while maintaining high answer accuracy.
Solution Approach 2:
The attention mechanism framework serves multiple functions simultaneously: it evaluates topic relevance, reasons, and causality through different attention weights. This multi-functional approach handles all three answer conditions within a unified system architecture, achieving high accuracy without proportionally increasing system complexity.
3Measurement precision
If the system evaluates all three conditions (topic relevance, reasons, causality) comprehensively, then the answer accuracy is improved, but the computational complexity increases
Solution Approach 1:
The patent performs preliminary actions by pre-defining the three evaluation aspects (topic, reason, causality) and their corresponding attention mechanisms. This preliminary structuring allows the system to efficiently evaluate all three conditions simultaneously during inference, rather than performing complex computations sequentially, thus improving accuracy while managing computational complexity.
Solution Approach 2:
The patent merges the evaluation of three separate conditions into a unified attention mechanism framework. By combining topic attention, reason attention, and causality attention into a single multi-functional attention system, the patent achieves comprehensive evaluation without the computational overhead of three separate independent systems.
Data Source
AI summary
A question answering device includes: a general word vector converter converting a question and an answer to semantic vectors in accordance with general context; a general sentence level CNN 214, in response to similarities of semantic vectors between words in question and answer and to strength of causality between the words, for weighting each semantic vector to calculate sentence level representations of the question and the answer; a general passage level CNN 218, in response to similarity between sentence level representations of question and answer, and to strength of relation of vectors in the sentence level representations viewed from causality, for weighting the sentence level representation to calculate a passage level representation for the question and answer passage; and a classifier determining whether or not an answer is a correct answer, based on the similarities between outputs from CNNs 214 and 218.