Machine Learning Domain Adaptation via Syntactic Vector Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing domain adaptation techniques for machine learning models often degrade in accuracy due to the inclusion of inappropriate training data from various subdomains, as they do not consider syntactic information of verbs and named entities when selecting training data, leading to poor performance in specific subdomains like biomedical subdomains.
Innovation Solution
The information processing apparatus identifies named entities and verbs in dependent relationships within sentences, vectorizes them, and selects training data based on similarity thresholds to retrain machine learning models, ensuring accurate domain adaptation by focusing on syntactic representations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If domain adaptation is performed using training data from various subdomains, then the model can be applied to multiple domains, but the accuracy degrades due to inclusion of inappropriate training data
Solution Approach 1:
The patent segments the training data selection process by identifying and extracting specific syntactic components (named entities and verbs in dependent relationships) from sentences. This segmentation allows the system to evaluate and select training data at a fine-grained syntactic level, ensuring that only appropriate sentences from various subdomains are selected for domain adaptation, thus maintaining both versatility and accuracy.
Solution Approach 2:
The patent applies local quality by focusing on specific syntactic features (named entities and their dependent verbs) rather than treating all training data uniformly. By vectorizing and comparing these local syntactic components, the system can selectively adapt to different domains while maintaining high accuracy, as each domain's unique syntactic characteristics are preserved and utilized appropriately.
2Ease of operation
If training data is selected without considering syntactic information, then the selection process is simpler, but the model performance in specific subdomains deteriorates
Solution Approach 1:
The patent replaces manual or rule-based syntactic analysis with automated vectorization and similarity computation. By converting named entities and verbs into vector representations and using computational similarity measures, the system automatically selects appropriate training data based on syntactic information, maintaining ease of operation while significantly improving subdomain performance through principled syntactic matching.
3Quantity of substance
If all sentences from target domain are used for retraining, then more training data is available, but inappropriate data reduces model accuracy
Solution Approach 1:
The patent extracts and selects only the appropriate sentences from the target domain training data by comparing their syntactic vectors against reference patterns. This extraction process removes inappropriate sentences while retaining valuable ones, ensuring that the model is retrained on a curated subset of high-quality training data that maintains both sufficient quantity and high accuracy.
Data Source
AI summary
A non-transitory computer-readable recording medium stores a program for causing a computer to execute a process, the process includes identifying first named entities and first verbs in a dependent relationship respectively with the first named entities from each of a plurality of sentences, vectorizing each of the plurality of sentences based on the first named entities and the first verbs, identifying, out of the plurality of sentences, one or more sentences each having a similarity greater than or equal to a threshold with respect to a specific sentence based on a plurality of vectors generated by the vectorizing, the similarity between two sentences being defined as a similarity between vectors generated by vectorizing the two sentences, and executing training of a machine learning model based on the one or more sentences.