Language-Model Taxonomy Expansion for Sparse Phrase Sets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for expanding a taxonomy are inaccurate in cases where the number of phrases is small or when search queries and user action logs are unavailable.
Innovation Solution
An information processing apparatus and method that utilize a language model to generate and associate new phrases with existing phrases in a taxonomy, enhancing accuracy through the use of a first sentence group and a language model for generating and expanding processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing taxonomy expansion techniques are used, then the expansion can be performed, but the accuracy is insufficient when the number of phrases is small or sentence group is limited
Solution Approach 1:
The patent introduces a language model as an intermediary component that processes the relationship between new phrases and existing taxonomy phrases. The language model serves as a mediator that computes similarity scores and determines appropriate parent-child relationships, enabling accurate taxonomy expansion even when the input data (sentence group) is limited. This intermediary mechanism resolves the contradiction by providing a robust computational framework that doesn't heavily depend on the quantity of input phrases.
Solution Approach 2:
The patent changes the parameter of phrase representation by using embedding vectors generated through the language model. Instead of relying on traditional text matching or surface-level features, the system transforms phrases into high-dimensional vector representations that capture semantic meaning. This parameter transformation enables accurate comparison and relationship determination even with a small number of phrases, thereby improving expansion accuracy without requiring large quantities of input data.
2Measurement precision
If existing taxonomy expansion techniques are used, then the expansion can be performed, but the accuracy is insufficient when search query and user action log are unavailable
Solution Approach 1:
The patent extracts and removes the dependency on external data sources (search queries and user action logs) from the taxonomy expansion process. By designing a system that operates solely on the basis of a sentence group and existing taxonomy, the invention extracts the essential expansion functionality while discarding the requirement for additional information sources. This extraction resolves the contradiction by making the system self-sufficient and accurate without needing search queries or user behavior data.
Solution Approach 2:
The patent implements a self-service mechanism where the language model and embedding vectors enable the system to perform taxonomy expansion using only its internal resources (the sentence group and existing taxonomy structure). The system serves itself by computing phrase similarities and determining relationships through the language model, without requiring external validation or supplementary data from search queries or user logs. This self-service capability maintains accuracy while eliminating the loss of information associated with unavailable external data sources.
3Measurement precision
If language model is used for generating and expanding phrases, then the accuracy of taxonomy expansion is improved, but the computational complexity increases
Solution Approach 1:
The patent segments the taxonomy expansion process into distinct computational stages: (1) generating embedding vectors for new phrases using the language model, (2) computing similarity scores between new phrases and existing taxonomy phrases, and (3) determining parent-child relationships based on similarity thresholds. This segmentation allows the system to apply the computationally intensive language model only where necessary (for embedding generation) while using simpler operations for comparison and relationship determination, thereby managing overall computational complexity while maintaining high accuracy.
Data Source
AI summary
In order to expand a taxonomy more accurately, an information processing apparatus (1) includes: an obtaining section (11) that obtains a taxonomy which indicates a relationship between a plurality of phrases; a generating section (12) that generates a new phrase which differs from each of the plurality of phrases, with reference to a first sentence group; and an expanding section (13) that expands the taxonomy by associating the new phrase with any of the plurality of phrases. One or both of the generating section (12) and the expanding section (13) use a language model.