Semantic Event Template Matching for Sparse-Corpus Intent Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Voice interaction systems struggle to accurately recognize user intentions due to limited corpora, leading to misunderstandings and poor interaction experiences as users express requests in various ways.
Innovation Solution
A semantic understanding method that preprocesses events, uses pre-trained semantic enhancement models for statement expansion, and calculates matching degrees to identify target event templates, incorporating weighted calculations and iterative learning to generate similar statements for improved recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a limited corpus is used for training the voice interaction system, then the system complexity is reduced, but the ability to accurately recognize user intentions deteriorates
Solution Approach 1:
The patent uses a semantic enhancement model to generate similar statements (copies) of existing corpus entries. When the model expands a statement, it creates multiple variant expressions that preserve the original meaning but use different wording, structures, or expressions. This copying approach allows the system to effectively increase corpus diversity without proportionally increasing the actual training data volume, thereby improving recognition accuracy while maintaining relatively controlled system complexity
Solution Approach 2:
The patent changes the semantic parameters of existing statements by adjusting wording, sentence structure, and expression patterns while maintaining the core meaning. The semantic enhancement model modifies statements by replacing words with synonyms, changing active voice to passive voice, or rephrasing sentences, thereby creating diverse training samples from limited original data and improving the system's ability to recognize various expressions of the same user intention
2Adaptability or versatility
If more diverse expressions are supported to adapt to user habits, then the adaptability improves, but the corpus requirements increase leading to system complexity increase
Solution Approach 1:
The patent performs preliminary semantic enhancement on the corpus during the training phase. By pre-generating similar statements and expanding the semantic coverage of the corpus in advance, the system prepares diverse expression patterns before actual use. This preliminary action ensures that when users express themselves in various ways, the system has already encountered similar patterns during training, thereby improving adaptability without requiring an excessively large original corpus
Solution Approach 2:
The semantic enhancement model serves multiple functions: it generates training data, creates test cases, and provides semantic generalization capabilities. By using the same model for statement expansion, similarity calculation, and template matching, the system achieves multi-functionality that allows diverse expression support without proportionally increasing system complexity, as the same computational infrastructure handles multiple tasks
Data Source
AI summary
A semantic understanding method includes calculating a first matching degree between an event to be processed and a preset event template, performing statement expansion on the event to be processed to obtain a similar event, calculating a second matching degree between the similar event and the preset event template, obtaining a first result and a second result according to the first matching degree and the second matching degree, and taking the preset event template of a larger one for semantic understanding.