Tweet sentiment quantification: An experimental re-evaluation.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Social Science

Background:

• Sentiment quantification estimates sentiment class prevalence in unlabeled text, crucial for Twitter data analysis.
• The 'classify and count' method is known to be suboptimal for accurate sentiment quantification.
• Previous systematic comparisons of tweet sentiment quantification methods have questionable reliability due to weak experimentation.

Purpose of the Study:

• To re-evaluate existing and modern sentiment quantification methods for tweets.
• To address the limitations of prior experimental protocols in sentiment quantification research.
• To provide a more reliable understanding of method performance under varying prevalence conditions.

Main Methods:

• Re-evaluation of quantification methods on existing tweet sentiment datasets.
• Implementation of a robust experimental protocol, including simulated prevalence shifts.
• Comparison of multiple quantification techniques, including those not previously assessed.

Main Results:

• Experimental results differ significantly from prior studies due to improved methodology.
• The robust protocol revealed distinct performance characteristics of various sentiment quantification methods.
• A more reliable understanding of the strengths and weaknesses of different approaches was achieved.

Conclusions:

• The reliability of sentiment quantification methods is highly dependent on experimental design.
• Modern and re-evaluated methods offer improved accuracy for estimating sentiment prevalence in tweets.
• This study provides a more solid foundation for selecting appropriate methods for tweet sentiment quantification.