Natural Language Processing Research

A new method of approaching sentiment classification is proposed where the likelihood of word embeddings to produce useful information from limited Twitter data is studied. The novelty of this work is in determining how short corpuses (taken from Twitter data) are polarized to multiple axes with respect to a subject, as opposed to using a single positive-negative sentiment axis to classify the text with respect to a subject. The unique methodology of this model focuses on deconstructing a short corpus (microblogging entry from Twitter) into key tokens, identifying the correct axis of the sentiment (the polarization axis) using cosine similarity, and then using this axis to generate polarization values to classify each selection of text into fine-tuned axis values. Results of this study show that a single axis may not be enough to express a sentiment. Various axes will have to be combined for better results. Results were measured in terms of classification accuracy, classification bias, and an axis score.