Semantic sentiment based LDA for detecting sentiments of short texts

Semantic sentiment based LDA for detecting sentiments of short texts

연구 분야: Strategies

학회: World Wide Web

This paper presents a novel topic modeling approach, Semantic Sentimental LDA (SSLDA), designed to improve sentiment detection in short texts by leveraging prior sentiment knowledge from reliable sources. Specifically, we utilize emojis extracted from a dataset of approximately 8 million tweets to refine sentiment classification. The methodology follows a structured 9-stage process, where sentiment-bearing words identified in earlier stages are progressively integrated into subsequent steps to enhance classification robustness. Rather than employing traditional supervised methods, our approach iteratively refines a sentiment lexicon, termed the Golden List, which distinguishes sentiment-positive and sentiment-negative words—including previously unrecognized terms used informally in microblogging contexts. Comparative analysis against existing sentiment lexicons demonstrates a higher rate of sentiment alignment, validating the effectiveness of SSLDA in addressing the challenges of informal text sentiment classification.