Learning word sense disambiguation in biomedical text with difference between training and test distributions

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Machine learning methods for word sense disambiguation (WSD) often face challenges due to distribution mismatches between training and testing datasets.
• Lexical diversity in natural language contributes to these distribution discrepancies, impacting model generalization.

Purpose of the Study:

• To address the performance degradation caused by distribution discordance in WSD.
• To propose and evaluate a novel approach, Support Vector Machines with Example-wise Weights (SVM-EW), for improving WSD.

Main Methods:

• The proposed SVM-EW method adjusts the training data distribution to better align with the test data distribution.
• Training examples are weighted based on their similarity to the entire set of test data.
• This approach aims to mitigate the impact of lexical space diversity.

Main Results:

• Experimental results demonstrate that the proposed method effectively recognizes and accounts for distribution changes between training and test data.
• SVM-EW significantly outperforms standard Support Vector Machines (SVMs) in WSD tasks when distribution shifts are present.

Conclusions:

• Weighting training examples according to their similarity to test data is a viable strategy to handle distribution mismatches in WSD.
• The SVM-EW method offers a robust solution for improving the accuracy and generalizability of WSD systems in diverse lexical environments.