Access Type

Open Access Thesis

Date of Award

January 2017

Degree Type

Thesis

Degree Name

M.S.

Department

Computer Science

First Advisor

Dongxiao Zhu

Abstract

Sentiment analysis is a process of learning the relationship between sentiment label

and text. The research value of sentiment analysis is two-fold: first, it has a wide range

of applications in many sectors and industries, e.g., the industry has flourished due to the

proliferation of commercial applications such as using sentiment analysis as an integrated

part of customer experience strategy. Second, it offers an array of new challenging problems

for research community such as word feature embedding and machine learning. Albeit earlier

methods such as Naïve Bayes (NB), Random Forest (RF), k-Nearest-Neighbours (kNN),

Support Vector Machine (SVM) and more recent methods such as Deep Learning (DL)

methods are effective, they are primarily designed for shorter or longer textual data thus

are not able to maintain a robust performance across a variety of text with diverse lengths.

In reality, some text is as abbreviated as one single word while others are so pleonastic

that are over thousands of words. Moreover, ad hoc combination of feature embedding and

learning methods makes it more difficult to choose the right approach for different types of

textual data. Undoubtedly an integrated feature embedding and sentiment analysis method

is desirable. In this thesis, we introduce multi-way FM as a new method for sentiment analysis

accounting for higher-order feature interaction. We demonstrate the performance and

flexibility of the FM method to other competing methods by tuning a single parameter to

accommodate both shorter Twitter and longer movie review documents.

Share

COinS