Extraction Device for Natural Language Evaluation Factor Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to accurately analyze character information and extract words indicating evaluation factors, such as 'positive' or 'negative', or suggestions, like requests, from customers due to sensitivity to orthographic variants in dependency analysis.
Innovation Solution
An extraction device and method that acquire a natural sentence, extract target words relevant to evaluation factors using a model learned from a second natural sentence, and output the extracted target words, thereby overcoming the sensitivity issues in dependency analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If dependency analysis through BERT is used to extract target words, then the extraction can capture contextual meaning, but the result becomes sensitive to orthographic variants making rule-based extraction difficult
Solution Approach 1:
The patent introduces an intermediary process between BERT dependency analysis and rule-based extraction. The system uses BERT to generate dependency relationships, then applies a learned model that maps these dependencies to extraction rules, serving as a mediator that bridges the gap between contextual understanding and rule-based reliability
Solution Approach 2:
The patent changes the parameters of the extraction system by transitioning from pure rule-based extraction to a hybrid approach using learned models. The model learns from training data to adapt extraction parameters dynamically, reducing sensitivity to orthographic variants while maintaining the benefits of contextual analysis
2Measurement precision
If a learned model is used to extract relevant words and target words, then extraction accuracy improves, but processing time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-training the extraction model on large datasets of natural sentences with annotated relevant words and target words. This pre-learning phase captures common patterns and relationships, enabling faster and more accurate extraction during actual use without requiring complex real-time computations
Data Source
AI summary
An extraction device 400 includes: an acquisition section 421 acquiring a first natural sentence input by a user; an extraction section 422 extracting at least a target word of a relevant word and the target word from the first natural sentence acquired by the acquisition section 421 using a model learnt to output the relevant word and the target word with a second natural sentence as an input, the relevant word being a word defining the relevancy between words included in the second natural sentence and the target word being a word serving as the target of the relevant word; and an output section 423 outputting the target word extracted by the extraction section 422.