| Summary: | 碩士 === 國立臺北科技大學 === 電機工程研究所 === 105 === Sentiment Analysis is important for organizations and individuals who want to ensure that their reputations, influence and products are well received by their target audience. This importance is more pronounced for politicians especially when they are running for elections and trying to gain support of their constituents. Social networks become a critical tool for both dissipating their ideas and analyzing the response and sentiments attracted by each post. There has been limited research on Sentiment Analysis using Facebook data as most researchers bias their research towards Twitter data because of its limit on tweet length and hashtags which simplify topic detection. Moreover, most of this research is rarely focused on the data itself but the primary focus becomes the accuracy of the prediction models being built and if at all the data is looked at its mostly limited to a specific domain. Some researchers have studied the effect of the data pre-processing methods on the accuracy of the sentiment analysis algorithms. These researches and others only consider one variable i.e. text in their analysis yet there are other variables that have an effect on the kind of response and sentiment attracted by Facebook posts. This thesis, therefore, attempts to consider other variables in the Facebook sentiment analysis process. We factor-in the time at which the post was made, the type of media posted and topic of discussion. In a nutshell, it checks for all the variables that determine the amount and type of reception a Facebook post will receive. The prior polarity score feature model is used to extract features in the text while the Latent Dirichlet allocation (LDA) model is used for topic detection and the sliding window algorithm and other statistical tools and tests are used to get to the results. To perform the experiments we used data from the two 2016 United States of America (USA) presidential candidates’ Facebook posts and a major news network in the United States of America (USA) i.e. Fox news spanning from 1 January 2016 to 31 December 2016. The results show that the sentiments value received by a Facebook post is not independent of the other variables in the post and basing all sentiment prediction algorithms solely on text features will give results that are void of some level of interesting knowledge. The results of this study can be employed by both organizations and individuals to guide their social media, Facebook in particular, activity to further increase the chances of product sales and influence by increasing the level of coverage and sentiment that their posts receive.
|
|---|
