Question Answering Hierarchy for Non-Straightforward Queries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic question answering systems lack the ability to process non-straightforward queries and learn subtle nuances, leading to reduced accuracy in answering questions that do not have a straightforward relationship with the answer.
Innovation Solution
The system employs an architecture that generates responses by learning cues and using an abstraction hierarchy to map questions into informative terms, allowing it to find answers even when they are not contained in a string with a straightforward relation, and automatically updates the hierarchy to improve question answering accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional query rewriting methods are used to find straightforward answer relationships, then simple questions can be answered accurately, but non-straightforward queries cannot be processed effectively
Solution Approach 1:
The system dynamically adapts its approach based on the question type. For straightforward questions, it uses direct query rewriting; for non-straightforward questions, it employs machine learning models to learn subtle relationships from training data, allowing the system to flexibly adjust its methodology to maintain accuracy across diverse query types
Solution Approach 2:
The system changes the parameters of its processing approach based on question characteristics. It uses different feature extraction methods, weighting schemes, and matching strategies depending on whether the question is straightforward or non-straightford, enabling it to handle both types effectively with optimized parameters for each category
2Measurement precision
If extensive knowledge engineering and handwritten patterns are used, then question answering accuracy improves, but system complexity and manual labor increase significantly
Solution Approach 1:
The system performs self-training by automatically learning from training question-answer pairs without requiring extensive manual knowledge engineering. The machine learning models autonomously extract features, learn relationships, and optimize parameters, reducing the need for hand-written patterns while maintaining or improving accuracy
Solution Approach 2:
The patent replaces manual knowledge engineering and handwritten pattern creation with automated machine learning systems. The mechanical process of manually crafting rules is substituted with computational learning processes that automatically discover patterns from data, reducing system complexity while preserving accuracy
3Adaptability or versatility
If conventional systems process each question of the same type identically, then processing is simple, but the system cannot learn subtle nuances from new questions
Solution Approach 1:
The system performs preliminary learning during a training phase using training question-answer pairs. This preliminary action allows the model to learn subtle nuances and relationships before actual question processing, enabling it to handle new questions with learned knowledge while maintaining relatively simple processing during deployment
Solution Approach 2:
The system uses feedback from training data to continuously improve its performance. By processing training questions and comparing results with known answers, the model learns from feedback and adjusts its parameters and features, enabling it to capture subtle nuances while keeping the operational process streamlined
Data Source
AI summary
Architecture for generating responses to search engine queries. Question/answer pairs are used to provide a predefined question hierarchy to facilitate a search. Informative terms and one or more questions are associated with each question node of the hierarchy. A question posed is mapped into the hierarchy to the most similar question class node, as determined by passages of the question and/or comparing the informative terms of the question and the question/answer node. Weighted informative terms associated with that node are then returned, and can be used by an automatic question answering system to more effectively find an answer to a question.