Sentiment analysis in tweets: an assessment study from classical to modern word representation models

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Media Analysis

Background:

• Social media generates vast user-generated text data daily, particularly on platforms like Twitter.
• Analyzing sentiment in tweets is crucial for decision-making but challenging due to informal and noisy language.
• Existing research often evaluates tweet sentiment analysis models on limited datasets, creating a gap in robust performance assessment.

Purpose of the Study:

• To comprehensively assess neural language models for sentiment classification on tweets.
• To evaluate the effectiveness of static and contextualized word representations for tweet sentiment analysis.
• To investigate model adaptation strategies for the specific linguistic style of social media data.

Main Methods:

• Utilized a diverse collection of 22 datasets spanning multiple domains for robust evaluation.
• Employed five distinct classification algorithms to analyze model performance.
• Evaluated both static and contextualized representations, including Transformer-based autoencoder models fine-tuned via masked language modeling.

Main Results:

• Demonstrated that Transformer-based models, particularly when adapted to the nuances of tweet language, show superior performance in sentiment classification.
• Highlighted the importance of dataset diversity in evaluating the generalizability of sentiment analysis models.
• Identified specific adaptation strategies that significantly improve model performance on noisy, short-form social media text.

Conclusions:

• Neural language models offer powerful capabilities for tweet sentiment analysis, but their effectiveness is highly dependent on model architecture and data-specific fine-tuning.
• A robust evaluation framework using multiple datasets and algorithms is essential for understanding model limitations and strengths.
• Future research should focus on developing and adapting models that inherently handle the unique characteristics of social media language for improved sentiment analysis accuracy.