Aspect-Based Unsupervised Negative Sentiment Analysis

Twitter is a social media platform where users post their opinions on various events, products, services and celebrities. Automated analysis of these public posts is useful for tapping into public opinion and sentiment. Identifying negative public sentiment assumes importance when national security issues are at stake or when critical analysis of a product or policy is required. In this paper, a method is introduced that classifies tweets based on their negative content, without any prior training. Specifically, an unsupervised negative sentiment analysis is presented using an aspect-based approach. Phrase and keyword selection criteria are devised after identifying fourteen valid combinations of part-of-speech tags listed in a prioritized order, that are defined as phrase patterns. A sliding text window is passed through each sentence of the tweet to detect the longest valid phrase pattern. The keyword indicating the aspect information is detected using a dependency parser. SentiWordNet lexicon is used for scoring the terms in the detected keyword and phrase combination. The scores are summed up for each sentence of the tweet and transformed nonlinearly by a modified sigmoid function whose output is in the range [−2, 2]; this value comes out to be negative for negative tweets. The utility of our method is proved by superior results as compared to the state of the art on the benchmark SemEval 2013 twitter dataset.