Enhancing Lexical Sentiment Analysis using LASSO Style Regularization

Abstract

In the current information age where expressing one’s opinions online requires but a few button presses, there is great interest in analyzing and predicting such emotional expression. Sentiment analysis is described as the study of how to quantify and predict such emotional expression by applying various analytical methods. This realm of study can broadly be separated into two domains: those which quantify sentiment using sets of features determined by humans, and approaches that utilize machine learning. An issue with the later approaches being that the features which describe sentiment within text are challenging to interpret. By combining VADER which is short for Valence Aware Dictionary for sEntiment Reasoning; a lexicon model with machine learning tools (simulated annealing) and k-fold cross validation we can improve the performance of VADER within and across context. To validate this modified VADER algorithm we contribute to the literature of sentiment analysis by sharing a dataset sourced from Steam; an online video game platform. The benefits of using Steam for training purposes is that it contains several unique properties from both social media and online web retailers such as Amazon. The results obtained from applying this modified VADER algorithm indicate that parameters need to be re-trained for each dataset/context. Furthermore that using statistical learning tools to estimate these parameters improves the performance of VADER within and across context. As an addendum we provide a general overview of the current state of sentiment analysis and apply BERT a Transformer-based neural network model to the collected Steam dataset. These results were then compared to both base VADER and modified VADER.