Feature engineering for drug name recognition in biomedical texts: feature conjunction and feature selection

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Bioinformatics

Background:

• Drug name recognition (DNR) is crucial for extracting drug information.
• Current machine learning methods for DNR often rely on insufficient singleton features.
• Combining features can lead to excessive features and noise, limiting performance.

Purpose of the Study:

• To explore the effectiveness of feature conjunction and feature selection for DNR.
• To address the limitations of singleton features in capturing comprehensive linguistic characteristics for DNR.
• To improve the performance of drug name recognition systems.

Main Methods:

• Selected 8 types of singleton features for DNR.
• Combined singleton features into conjunction features using two distinct methods.
• Employed Chi-square, mutual information, and information gain for feature selection.
• Evaluated the enhanced DNR system on the DDIExtraction 2013 challenge dataset.

Main Results:

• Feature conjunction and selection demonstrably improve DNR system performance.
• The proposed method achieves improved results with a manageable number of features.
• The enhanced DNR system significantly surpasses the top-performing system from the DDIExtraction 2013 challenge.

Conclusions:

• Feature conjunction and selection are effective strategies for enhancing DNR.
• The developed approach offers a more robust method for drug name recognition.
• This work advances the field of drug information extraction through improved feature engineering.