Bachelor's Dissertation: Sentiment Analysis

Exploring geographical bias in sentiment analysis as a result of stylistic word choices (using tweets from London boroughs as our environment)

Abstract

In recent years sentiment analysis has emerged as an important research topic because of its ability to summarise public opinions. The most prevalent domain used is Twitter, which provides access to an immeasurable archive of opinionated text. However, research on the fairness of these algorithms has stalled. This is due to the absence of access to the protected characteristics of Twitter users.

This dissertation defines a new form of theoretical bias in these algorithms called geographical bias and explores the existence of this novel concept across an urban area. This empirical investigation reveals that the sentiment analysis algorithms did exhibit significant geographical bias in the environment defined. However, the limitations of the methodology restrict the extension of this environment to the real world. Another finding is that sentiment scores exhibited location-based clustering of insignificantly different distributions. This dissertation concludes with a critical evaluation of the methods used and an outline for further work building on these new concepts.