In this thesis, we explore several specific aspects of Twitter sentiment analysis. Our system for sentiment analysis is based on machine learning and text mining techniques, such as the bag-of-words representation of texts and support vector machine classifier. We employ our system to analyze a stream of short messages (tweets) about financial markets, specifically about stock trading, in the time span of two years. We classify each message into positive, negative, or neutral class, which represent the sentiment or stance towards the stock mentioned in the message. The term sentiment in our case thus denotes the stance of the author (speaker) and in the case of positive or negative class represents the author’s leaning towards buying or selling the stock. To build the classification model, we employ a relatively large gold standard which consists of approximately a half million tweets hand-labeled by the domain experts.
For the purpose of this analysis, we developed an evaluation platform and a methodology that allow us, by conducting a series of experiments, to answer various questions which arise when applying sentiment analysis in industrial settings. In the evaluation processes, we take the temporal nature of the data into account and thus enable continuous monitoring of performance of live systems.
The results of the analysis reveal (i) the most appropriate classification algorithm, (ii) the optimal size of the labeled data and subsampling method, (iii) the relationship between the classifier performance and the time lag from the training data, and (iv) the effect of duplicated tweets (e.g., retweets), and (v) the behavior of the employed classification method in the uncertainty area near the hyper-plane of support vector machine classifier.
|
