Feature engineering for sentiment analysis in e-health forums

Area of Science:

• Computational linguistics
• Health informatics
• Social media analysis

Background:

• Health-related social media generates vast amounts of data.
• Accessing and analyzing this information is challenging.
• Existing opinion mining methods often use traditional machine learning and bag-of-words approaches.

Purpose of the Study:

• To evaluate various features for classifying patient-authored texts for polarity.
• To investigate the significance of "polar facts" in health information analysis.
• To improve the accuracy of sentiment analysis in online health discussions.

Main Methods:

• Annotated over 3500 posts from online health forums (breast cancer, Crohn's, allergies).
• Classified sentences as "experience," "fact," or "opinion," and "positive," "negative," or "neutral."
• Trained machine learning models comparing bag-of-words with word embeddings and UMLS concept embeddings.

Main Results:

• Word embeddings achieved high accuracy (≈70%) in polarity prediction, outperforming bag-of-words.
• Objective "polar facts" are as frequent as subjective opinions/experiences in health data.
• Polar facts are crucial for accurate health information polarity classification.

Conclusions:

• Word embeddings offer a superior method for health-related text polarity classification.
• The inclusion of objective polar facts is essential for comprehensive sentiment analysis of health information.
• This research facilitates better access to relevant health data from social media.