Abstract
Sentiment analysis or opinion mining seeks to automate the detection of subjectiveness underlying a text. It is essential for many applications, such as political campaign, online marketing, and products reviews. In the past decade, a line of researchers has studied the problem of automating sentiment analysis. The research in this area falls mainly into two directions: 1) classification models that optimize algorithms and features to predict the polarity of a text, and 2) lexiconbased models that utilize lexicons and rule-based approaches to determine the sentiment of a given text. In this paper, we present a system that constructs lexicon by utilizing both dictionary and context-based approaches and 2,25 million tweets collected from Twitter. The dictionary-based approach finds semantically associated keywords; however, this approach may not detect contextually related keywords. Thus, we apply context-based expansion utilizing neural networks and word-embeddings to find both syntactically and semantically related keywords within corpora without drift. Experimental results suggest that the proposed model yields better lexicons and outperforms the baseline model.