Inducing Domain-Specific Sentiment Lexicons from Unlabeled Corpora

Area of Science:

• Computational Social Science
• Natural Language Processing
• Computational Linguistics

Background:

• Sentiment analysis accuracy is domain-dependent.
• Existing sentiment lexicons often lack domain specificity.
• Computational social science research necessitates tailored sentiment analysis tools.

Purpose of the Study:

• To develop a computational framework for inducing domain-specific sentiment lexicons.
• To create accurate sentiment lexicons using only corpus data.
• To release historical and community-specific sentiment lexicons for public use.

Main Methods:

• Combining domain-specific word embeddings with a label propagation framework.
• Utilizing small sets of seed words to initiate lexicon induction.
• Employing a purely corpus-based approach, avoiding reliance on external curated resources like WordNet.

Main Results:

• Achieved state-of-the-art performance in inducing domain-specific sentiment lexicons.
• Demonstrated superior performance compared to methods using hand-curated resources.
• Induced historical sentiment lexicons for 150 years of English and community-specific lexicons for 250 Reddit communities.

Conclusions:

• The developed framework effectively generates accurate, domain-specific sentiment lexicons.
• Over 5% of English sentiment-bearing words have changed polarity in the last 150 years.
• Sentiment varies significantly across different online communities, underscoring the need for localized lexicons.