Query Expansion Model for FAQ Search Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current FAQ search systems face challenges in accurately retrieving relevant search results due to the reliance on comparing input queries primarily with first texts, leading to insufficient search accuracy when keywords do not match those in the second texts, and conventional query expansion methods fail to generate suitable keywords for the second texts.
Innovation Solution
A model generation device and text search device that learn a query expansion model using text pairs, generating an expanded query by filtering and sorting keywords based on importance, and using a neural network to integrate search scores from both first and second texts for improved relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional query expansion methods are used to expand keywords in the input query, then the number of keywords is increased to enable matching with a wider variety of documents, but the expanded keywords are not suitable for searching in the second texts (answers)
Solution Approach 1:
The patent segments the search process into two distinct parts: searching first texts (questions) with original keywords, and searching second texts (answers) with expanded keywords. This segmentation allows each search to use appropriately tailored keywords, resolving the contradiction between expanding keyword coverage and maintaining search precision for answer texts.
Solution Approach 2:
The patent introduces an intermediary model (the learned query expansion model) that generates expanded keywords specifically tailored for searching answer texts. This intermediary translates the input query into a form that is optimized for answering, rather than using conventional expansion methods that may generate irrelevant keywords for the answer context.
2Productivity
If only the first texts (questions) are used for search comparison, then the search process is simple and fast, but the search accuracy decreases when keywords do not match those in the second texts (answers)
Solution Approach 1:
The patent performs preliminary action by generating expanded keywords before searching the second texts. This allows the system to prepare appropriate search terms in advance, enabling accurate matching with answer texts without compromising search efficiency. The expansion model is pre-trained and can quickly generate relevant keywords during the search process.
Solution Approach 2:
The patent adds another dimension to the search process by introducing a second search path that operates on expanded keywords specifically for answer texts. This dimensional expansion allows the system to simultaneously maintain simple comparison with original questions while also performing enriched comparison with expanded keywords for answers, thereby improving accuracy without sacrificing efficiency.
3Adaptability or versatility
If a large number of FAQs and response candidates are stored in the database, then the system can provide more comprehensive answers, but the difficulty of selecting the most appropriate candidate increases
Solution Approach 1:
The patent applies local quality by using different search strategies for different parts of the data: original keyword matching for questions and learned expanded keyword matching for answers. This localized approach to different data types improves the overall effectiveness of candidate selection from the large database, making it easier to identify relevant answers among many candidates.
Data Source
AI summary
Taking as input a group of text pairs for learning in which each pair is constituted with a first text for learning and a second text for learning that serves as an answer when a question is made with the first text for learning, a query expansion model is learned so as to generate a text serving as an expanded query for a text serving as a query.