Opinion Mining Using Enriched Joint Sentiment-Topic Model

Abstract

Sentiment analysis has the potential to significantly impact several fields, such as trade, politics, and opinion extraction. Topic modeling is an intriguing concept used in emotion detection. Latent Dirichlet Allocation is an important algorithm in this subject. It investigates the semantic associations between terms in a text document and takes into account the influence of a subject on a word. Joint Sentiment-Topic model is a framework based on Latent Dirichlet Allocation method that investigates the influence of subjects and emotions on words. The emotion parameter is insufficient, and additional factors may be valuable in performance enhancement. This study presents two novel topic models that extend and improve Joint Sentiment-Topic model through a new parameter (the author’s view). The proposed methods care about the author’s inherent characteristics, which is the most important factor in writing a comment. The proposed models consider the effect of the author’s view on words in a text document. The author’s view means that the author creates an opinion in his mind about a product/thing before selecting the words for expressing the opinion. The new parameter has an immense effect on model accuracy regarding evaluation results. The first proposed method is author’s View-based Joint Sentiment-Topic model for Multi-domain. According to the evaluation results, the highest accuracy value in the first method is equal to 85%. It also has a lower perplexity value than other methods. The second proposed method is Author’s View-based Joint Sentiment-Topic model for Single-domain. According to the evaluation results, it achieves the highest accuracy with 95%. The proposed methods perform better than baseline methods with different topic number settings, especially the second method with 95% accuracy. The second method is a version of the first one, which outperforms baseline methods in terms of accuracy. These results demonstrate that the parameter of the author’s view improves sentiment classification at the document level. While not requiring labeled data, the proposed methods are more accurate than discriminative models such as Support Vector Machine (SVM) and logistic regression, based on the evaluation section’s outcomes. The proposed methods are simple with a low number of parameters. While providing a broad perception of connections between different words in documents of a single collection (single-domain) or multiple collections (multi-domain), the proposed methods have prepared solutions for two different situations (single-domain and multi-domain). The first proposed method is suitable for multi-domain datasets, but the second proposed method is suitable for single-domain datasets. While detecting emotion at the document level, the proposed models improve evaluation results compared to the baseline models. Eight datasets with different sizes have been used in implementations. For evaluations, this study uses sentiment analysis at the document level, perplexity, and topic coherency. Also, to see if the outcomes of the suggested models are statistically different from those of other algorithms, the Friedman test, a statistical analysis, is employed.