Non-Factoid QA System SVM Feature Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current question-answering systems for non-factoid questions, such as why-questions and how-to-questions, have lower accuracy compared to factoid question-answering systems and require improvement to effectively handle complex queries that involve reasoning and inference.
Innovation Solution
A non-factoid question-answering system that utilizes morphological analysis, semantic class conversion, and evaluation polarity features to generate and rank answer candidates, incorporating supervised machine learning with SVMs to improve the accuracy of answer selection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional term frequency and document frequency based scoring is used for answer selection, then the system can process non-factoid questions, but the accuracy of answer selection remains low
Solution Approach 1:
The system segments the answer selection process into multiple independent feature components: morphological features, syntactic features, semantic features, and evaluation polarity features. Each feature type is generated and scored separately, then combined to produce the final answer ranking. This segmentation allows the system to capture multiple aspects of question-answer relevance without overwhelming complexity in any single feature generator.
Solution Approach 2:
The system introduces evaluation polarity as a new dimension for feature generation, classifying phrases into positive, negative, and neutral categories. This adds a semantic orientation dimension to the traditional term frequency-based features, enabling the system to capture the evaluative nature of non-factoid questions and answers, thereby improving answer selection accuracy beyond what conventional methods achieve.
2Adaptability or versatility
If simple term matching and frequency-based scoring is used, then the system structure remains simple, but it cannot effectively handle complex queries involving reasoning and inference
Solution Approach 1:
The system performs preliminary morphological analysis and syntactic parsing on both questions and answer candidates before scoring. This preliminary processing extracts structural and semantic features that are essential for handling complex queries involving reasoning and inference, enabling the system to go beyond simple term matching and achieve better answer accuracy.
Solution Approach 2:
The system combines multiple types of features (morphological, syntactic, semantic, and evaluation polarity) into a composite scoring mechanism. This composite approach integrates various linguistic analyses to handle the complexity of non-factoid questions, allowing the system to effectively process queries involving reasoning and inference while maintaining high answer accuracy.
3Measurement precision
If morphological analysis and syntactic parsing are performed on all answer candidates, then answer accuracy improves, but processing time increases
Solution Approach 1:
The system segments the feature generation process into modular components that can be independently applied. Morphological analysis and syntactic parsing are performed only on answer candidates that pass initial filtering based on simpler features, rather than on all candidates. This segmentation reduces the overall processing time while maintaining high evaluation accuracy for the most promising candidates.
Solution Approach 2:
The system applies full morphological and syntactic analysis only to a subset of answer candidates that show promise based on preliminary scoring using simpler features. This partial application of complex analysis reduces processing time while still achieving high accuracy by focusing computational resources on the most relevant candidates.
Data Source
AI summary
In order to provide a non-factoid question answering system with improved precision, the question answering system (160) includes: a candidate retrieving unit (222), responsive to a question, extracting answer candidates from a corpus storage (178); a feature vector generating unit (232) for generating features from combinations of a question with each of the answer candidates; SVMs (176) trained to calculate a score of how correct a combination of the question with an answer candidate is, upon receiving the feature vector therefor; and an answer ranker unit (234) outputting the answer candidate with the highest calculated score as the answer. The features are generated on the basis of the results of morphological analysis and parsing of the question, a phrase in the question evaluated as being positive or negative as well as its polarity, and the semantic classes of nouns in the features.