Explainable Text Similarity for Robust Automated Q&A Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated Q&A systems struggle to accurately assess the similarity between customer queries and pre-stored answers due to variations in wording, typographical errors, and grammatical alterations, leading to ineffective retrieval of relevant information.
Innovation Solution
A hierarchical explainable (HE) similarity approach that assesses similarity at multiple levels, including phrase, word, and character levels, using a hierarchical explainable feature (HEF) vector to generate a HE similarity score with explainable reasons for the relationship between texts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional scalar similarity metrics are used to compare customer queries with pre-stored answers, then the system is simple to implement, but the accuracy of answer retrieval deteriorates due to variations in wording, typographical errors, and grammatical alterations
Solution Approach 1:
The patent segments text similarity assessment into multiple hierarchical levels: character level, word level, phrase level, and sentence level. Each level extracts specific features independently, allowing the system to capture nuanced linguistic variations at different granularities. This segmentation enables robust similarity measurement despite wording variations, typographical errors, and grammatical alterations while maintaining structured processing.
Solution Approach 2:
The patent transitions from traditional scalar similarity metrics to multi-dimensional hierarchical feature vectors. Instead of computing a single similarity score, the system generates HEF vectors containing features across multiple dimensions (character n-grams, word embeddings, phrase structures, syntactic patterns). This dimensional expansion allows comprehensive text comparison that accounts for various types of linguistic variations simultaneously.
2Reliability
If hierarchical explainable feature vectors are used to assess text similarity at multiple levels, then the accuracy of answer retrieval improves by capturing nuanced linguistic features, but the computational complexity increases
Solution Approach 1:
The patent performs preliminary text processing to generate hierarchical features before similarity computation. Character n-grams, word embeddings, and phrase extractions are pre-computed and stored as HEF vectors. This preliminary action organizes text data into structured representations that facilitate efficient comparison, reducing the computational burden during actual query processing while maintaining high retrieval reliability.
Solution Approach 2:
The hierarchical feature vectors serve as intermediaries between raw text input and similarity assessment. Instead of directly comparing customer queries with pre-stored answers, the system transforms both into HEF vectors that capture linguistic features at multiple levels. This intermediary representation abstracts away surface-level variations (typos, grammar) while preserving semantic meaning, enabling reliable similarity measurement with optimized computation.
Data Source
AI summary
The present teaching relates to a hierarchical and explainable (HE) similarity and use thereof. A target text is identified based on a source text. A HE similarity characterizes the similarity between the source and target texts in terms of multiple assessment categories and is computed based on source and target phrases generated via linguistic features. A HE feature vector is constructed with similarity scores at phrase, word, and character levels. The HE similarity is computed based on the HE feature vector and used to determine whether the target text related to the source text. The HE similarity is used to determine whether the target text relates to the source text.