Active Learning for Turkish Sentiment Analysis

Abstract

Sentiment analysis/classification is a widely studied problem of natural language processing and data mining. With the availability of social media, there are a lot of data but it's hard to find a labeled training set because of its high cost. The goal of active learning is to get a better or same performance with fewer training data. In this work, the feasibility of active learning scheme for Turkish sentiment analysis is investigated. As a result, the same performance with full training set could be obtained with only half of the training set selected by active learning. Moreover, the affects of different clustering algorithms used at the initial set selection are investigated.