Twitter sentiment analysis for marketing research

Abstract

With the popularity of social media networks, Natural Language Processing faced new challenges because of the dynamic nature of the data found in these networks. This thesis focuses on sentiment analysis and attempts to classify tweets based on their sentiment for a marketing application which summarizes the public's view of products. We propose a real time web application, Twitter MAT (.Marketing Analysis Tool) which lets the user search for any product or query any keyword and provides a rendition of the results found after sentiment analysis is performed. A timeline is also provided to visualise how the opinion about the topic or product chosen by the user has evolved over time depending on the sentiment of the public's view. Twitter MAT uses the Named Entity Recognition technique to perform sentiment analysis and uses other features such as discourse analysis to detect sentiment polarity and its strength. These techniques were proven to be effective based on the results obtained in the evaluation of the sentiment analysis. The accuracy obtained for the classification of tweets based on their sentiment and is 75% when compared to manually annotated datasets and the results of the surveys conducted. Moreover, when compared to datasets which focus on sentiment polarity strength, we reached a 100% agreement.