Applying active learning to supervised word sense disambiguation in MEDLINE

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Supervised word sense disambiguation (WSD) methods often require extensive annotated data.
• Reducing the annotation effort is crucial for developing practical WSD systems.

Purpose of the Study:

• To evaluate the integration of active learning (AL) strategies with supervised WSD.
• To determine if AL can decrease the number of annotated samples needed without compromising model quality.

Main Methods:

• Support vector machine (SVM) classifiers were used for disambiguating terms in the MSH WSD collection.
• Three uncertainty sampling-based active learning algorithms were compared against a passive learner (random sampling).
• Learning curves and area under the learning curve (ALC) were used for performance evaluation.

Main Results:

• Active learners (ALs) outperformed the passive learner (PL) in 89.8% of WSD tasks.
• ALs achieved 90% accuracy with 24 samples, compared to 38 samples for the PL, a 37% reduction in annotation effort.
• Analysis identified factors influencing AL performance, including early model quality and WSD task difficulty.

Conclusions:

• Integrating active learning with supervised WSD methods is effective.
• Active learning strategies can significantly reduce annotation costs for WSD.
• This integration also leads to improved disambiguation model performance.