Two-Stage Entity Classification for Polysemy Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating the meaning of terms with polysemy in machine learning are inaccurate when there are many entities to be candidates, as they rely solely on similarity in term features without considering the broader context of writing.
Innovation Solution
A classification device and method that utilize two models: one for term features and another for writing features, determining linking based on similarities in both, improving accuracy by considering the context in which terms appear.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If one-stage classification using term feature similarity is used, then the process is simple, but accuracy deteriorates when there are many entity candidates
Solution Approach 1:
The classification process is divided into two distinct stages: first, term-based classification to identify candidate entities similar to the input term; second, writing-based classification to select the most appropriate entity from candidates by comparing the input writing with reference writings. This segmentation allows each stage to focus on specific features, improving overall accuracy while maintaining manageable complexity.
Solution Approach 2:
The patent introduces an intermediary step of generating candidate entities based on term similarity before performing the final classification. This intermediary stage acts as a bridge between the simple term-based approach and the more accurate writing-based approach, filtering the search space to reduce complexity in the final classification stage.
2Use of energy by moving object
If only term feature similarity is used for classification, then the method is computationally efficient, but accuracy deteriorates due to polysemy and context ambiguity
Solution Approach 1:
The classification task is segmented into two phases: a computationally efficient term-based phase that quickly identifies candidate entities using feature similarity, and a more intensive writing-based phase that resolves polysemy and context ambiguity by comparing full writing contexts. This segmentation maintains computational efficiency in the first phase while improving accuracy in the second phase.
Solution Approach 2:
The patent performs preliminary classification using term feature similarity to generate a narrowed-down set of candidate entities before conducting the more computationally intensive writing-based classification. This preliminary action reduces the search space, making the subsequent accurate classification more efficient while maintaining high accuracy in resolving polysemy.
Data Source
AI summary
Even when there are many candidates for a linked entity, accuracy of estimating an entity related to a term that appears on given writing is improved. A classification device stores a first model generated by using a corpus including pieces of first writing related to predetermined entities, and a second model generated by using a corpus including the pieces of first writing. Features of terms included in the first writing are acquired from the first model. Features of the pieces of first writing are acquired from the second model. The classification device causes the first model to output first information indicating entities indicated by terms, which has similarity in a feature to a second term included in second writing, among first terms indicating the predetermined entities. The classification device causes the second model to output second information indicating entities related to pieces of writing, which has similarity in a feature to the second writing, among the pieces of first writing. The classification device determines linking of the second term based on the first information and the second information.