Opinion Mining From Social Media Short Texts: Does Collective Intelligence Beat Deep Learning?

Area of Science:

• Computational linguistics
• Social media analytics
• Artificial intelligence

Background:

• Big data and AI enable real-time information mining from vast user-generated content.
• Social media marketing increasingly relies on opinion mining, with platforms like Google and Facebook developing proprietary tools.
• Current research predominantly focuses on Twitter sentiment analysis due to API accessibility, overlooking more popular platforms like Facebook and Instagram.

Purpose of the Study:

• To compare low-level features for sentiment analysis, including deep learning methods (fastText, Doc2Vec) and crowd-sourced keywords (crowd lexicon).
• To evaluate the effectiveness of these features and various machine learning models for sentiment analysis on tweets and Facebook comments.
• To investigate the impact of human annotation on developing effective AI tools for opinion mining.

Main Methods:

• Feature extraction using deep learning models (fastText, Doc2Vec) and a crowdsourcing platform for keyword generation (crowd lexicon).
• Sentiment analysis model development using various machine learning algorithms.
• Comparative analysis of feature types and machine learning methods on tweet and Facebook comment datasets.

Main Results:

• Deep learning features and crowd lexicons show varying effectiveness in sentiment analysis.
• Machine learning models trained on human-annotated data demonstrate superior performance.
• The study highlights the value of the 'learning by example' paradigm through human annotation.

Conclusions:

• Human annotation of even a small data portion is crucial for developing effective AI-driven sentiment analysis tools.
• Integrating crowd-sourced insights alongside deep learning features enhances opinion mining capabilities.
• Future research should explore opinion mining on more popular social media platforms beyond Twitter.