An Exploratory Study on Latent-Dirichlet Allocation Models for Aspect Identification on Short Sentences

Abstract

This paper reports an exploratory study conducted to investigate the performance of topic modelling algorithms in aspect identification. Aspect identification is an important step in aspect-based sentiment analysis. Latent-Dirichlet Allocation model serves as the baseline of topic models in the experiments. One of the variations of LDA, namely Phrase-LDA was experimented to benchmark its performance against the original LDA. Although it was reported that PLDA performs better compared to LDA in aspect-based sentiment analysis, our experimental results indicate that LDA works better on dataset with short sentences. A new PLDA model was also proposed by using different types of dependencies to extract the phrases.