LJST: A Semi-supervised Joint Sentiment-Topic Model for Short Texts

Several methods on simultaneous detection of sentiment and topics have been proposed to obtain subjective information such as opinion, attitude and feelings expressed in texts. Most of the techniques fail to produce desired results for short texts. In this paper, we propose LJST, a labeled joint sentiment-topic model particularly for short texts. It uses a probabilistic framework based on latent Dirichlet allocation. LJST is semi-supervised—it predicts the sentiment values for unlabeled texts in presence of a partially labeled texts with sentiment values. To address the sparsity problem in short text, we modify LJST and introduce Bi-LJST, which uses bi-terms (all possible pairs of words in a document) in place of unigrams for learning the topics by directly generating word co-occurrence patterns in each text and expressing the topics in terms of these patterns. Specifically, we have proposed a semi-supervised approach of extracting joint sentiment-topic model for short texts by incorporating bi-terms. Extensive experiments on three real-world datasets show that our methods perform consistently better than three other baselines in terms of document-level and topic-level sentiment prediction, and topic discovery—LJST using bi-term models outperforms the best baseline by producing 12% lower RMSE for document-level sentiment prediction and 6% higher F1 score for topic-sentiment prediction.