MSc Defence – Daniel Stantic

MSc candidate Daniel Stantic will defend his thesis "A Unified Probabilistic Model for Aspect-Level Sentiment Analysis" on March 9, 2016, at 2:30pm in Reynolds 219.

Title

A Unified Probabilistic Model for Aspect-Level Sentiment Analysis

Abstract

With the increasing availability of online opinionated text, research into sentiment analysis has seen an explosive amount of interest. In particular, aspect-level sentiment analysis aims to delve deeper into that content and discover how much people like specific aspects of products, services and other entities. Modern approaches use probabilistic models to tackle this challenge. In this thesis, we develop a new model based on POSLDA, a topic classifier that incorporates syntax modelling for better performance. POSLDA separates semantic words from purely functional words and restricts its topic modelling on the semantic words. We take this a step further by modelling the probability of a semantic word expressing sentiment based on its part-of-speech class and then modelling its sentiment if it is a sentiment word. We restructure the popular approach of topic-sentiment distributions within documents and add a few novel heuristic improvements. Our experiments demonstrate that our model produces results competitive to the state of the art systems. In addition to the model, we develop a multi-threaded version of the popular Gibbs sampling algorithm that can perform inference over 1000 times faster than the traditional implementation while preserving the quality of the results.

Advisor: Fei Song