Glossary Term Mapping Using Context-Enriched Semantic Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods using Bidirectional Encoder Representations from Transformers (BERT) struggle with the task of matching descriptive names in glossaries due to their brevity and generality, leading to a lack of context and increased computational load.
Innovation Solution
Employing Prompt-based Contrastive Learning for Sentence Embeddings (PromCSE) and Term Frequency-Inverse Document Frequency (TF-IDF) models to enrich descriptive names with relevant sentences from external knowledge bases, such as knowledge graphs, to enhance context and improve semantic text similarity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If conventional BERT methods are used for matching descriptive names in glossaries, then semantic understanding capability is provided, but the brevity and generality of descriptive names lead to lack of context and increased computational load
Solution Approach 1:
The patent segments the context enrichment process into distinct stages: first extracting candidate sentences from knowledge bases using TF-IDF, then filtering and selecting the most relevant sentences using PromCSE semantic similarity scoring, and finally mapping selected sentences to glossary terms. This segmented approach processes only necessary portions of data at each stage, reducing overall computational load while maintaining context quality.
Solution Approach 2:
The patent introduces PromCSE sentence embeddings as an intermediary mechanism between TF-IDF extracted sentences and the final glossary matching. This intermediary semantic similarity scoring system filters and ranks candidate sentences, selecting only those most relevant to each descriptive name, thereby reducing information loss without requiring computational processing of all candidate sentences.
2Ease of operation
If descriptive names in glossaries are kept brief and general, then ease of use and memorability improve, but context information is insufficient for accurate matching
Solution Approach 1:
The patent adds a new dimension of context by retrieving and associating external knowledge base sentences with brief descriptive names. Instead of expanding the descriptive names themselves, the system creates a multi-dimensional structure where each brief name is linked to multiple candidate sentences from knowledge bases, providing rich context while preserving the simplicity and ease of use of the original brief names.
Data Source
AI summary
A computer-implemented method (CIM), according to one embodiment, includes causing a first search to be performed on a first knowledge base for a first descriptive name, and extracting sentences from results of the first search. The method further includes running at least one predetermined deep learning model on the results of the first search for determining similarity scores for each of the extracted sentences, where each of the similarity scores defines a similarity score for an associated one of the extracted sentences and a first glossary. A first of the determined similarity scores is used to map terms of a second glossary to terms of the first glossary to enhance context provided in search results generated using the glossaries.