TAM: Targeted Analysis Model With Reinforcement Learning on Short Texts.

Area of Science:

• Social Media Analytics
• Natural Language Processing
• Machine Learning

Background:

• Social media mining is crucial for applications like trend identification and marketing.
• Existing topic modeling struggles with focused analysis of short documents due to sparse data and error propagation.
• Extracting specific topics from large volumes of short texts remains a challenge.

Purpose of the Study:

• To propose a targeted analysis model (TAM) using reinforcement learning (RL) for effective topic extraction from social media.
• To address limitations in current models for focused topic analysis in short documents.
• To enable fine-grained topic generation within specific domains.

Main Methods:

• Developed a targeted analysis model (TAM) integrating reinforcement learning (RL).
• Designed a novel reward function for RL to mitigate clustering errors from Gibbs sampling.
• Employed policy search combined with the Gibbs EM algorithm for parameter estimation.
• Utilized F1 score and normalized mutual information-F1 for evaluating clustering and topic generation.

Main Results:

• TAM demonstrates superior performance compared to state-of-the-art models in focused topic analysis.
• Achieved an average improvement of 25.7% in F1 score for binary clustering.
• Successfully extracts specific topics and generates fine-grained topic distributions from short texts.

Conclusions:

• The proposed TAM effectively addresses the challenges of targeted topic modeling in social media.
• Reinforcement learning integration significantly enhances the accuracy and robustness of topic extraction.
• TAM offers a promising approach for in-depth analysis of user-generated content.