Semi-automatic sentiment analysis based on topic modeling

Abstract

Sentiment is an important feature of natural language. It is used to understand semantic of texts and opinion of people. There are many practical applications, which require to extract sentiment from texts: advertising analytics, interactive chat bots, opinion mining. Today, different supervised techniques are used to extract sentiment from texts which require large manually labeled datasets that are expensive and time consuming to build. Moreover, such datasets should cover vocabularies and patterns of use of different contexts. Additionally, the efficiency of supervised methods trained on a well-written texts can dramatically decrease on users’ texts from social media due to typos, slang, short length of sentences.To solve these problems and to reduce human involvement, we propose semi-supervised sentiment analysis method based on topic modeling with Additive Regularization. To evaluate the efficiency of this method we applied it to several open-source datasets for which sentiment labels are available. The study shows promising results in terms of f1-score with minimal human involvement.