Comparative analysis of Active Learning strategies in Twitter domain

Official URL: http://risc01.sabanciuniv.edu/record=b1620704 (Table of Contents)

Since its launch in the year 2006, Twitter has been One of the most popular social media platforms where users are free to share opinions, ideas and feelings. Latest statistics tell us that nearly 350,000 tweets are being posted every minute On Twitter. Also twitter is the first place to track the response to any important incident or events in the world. For this reason, Twitter has attracted the researchers from many fields, including Sentiment Analysis which deals with opinion mining from text. Twitter data is rich in containing the sentiments but is inherent with the problem of being very informal and unstructured which makes it very difficult to convert this data information. Labeling this large amount of data build classifiers for supervised learning is next to impossible. So we make use of Active Learning which is a subfield of Machine Learning and concerns with the selection of most informative instances to train the classifiers thus saving labeling efforts. This thesis deals with the comparative analysis of selected Active learning sampling strategies with twitter domain. The results show Uncertainty Sampling beats Random Satnpling and (Query by Committee consistently An analysis of agreelllent levels among annotators for twitter data has also been presented.