Design and Development of Minimal Supervision Models for Aspect Level Sentiment Analysis

Abstract

Sentiment analysis is one of the prevalent and hotspot research areas. Sentiment analysis research has expanded laterally and has its applications in many other research areas. As the applications of sentiment analysis grew, one of the most fine-grained form in demand is Aspect level sentiment analysis, the field which identifies the product features/aspects that are talked about in the review and hence map a sentiment to each aspect. There has been active research in the field of sentiment analysis for a couple of decades where the task of aspect term extraction which identifies the product features/aspects in the review has been found more challenging. The large volume of unlabeled data definitely encourages the need for more unsupervised models to be experimented in the field. Supervised models are definitely outperforming unsupervised models but are constrained by the availability of labelled data for the fine grained task. There is a huge cost, time and effort involved in creating voluminous and qualitative labelled datasets for training data required by supervised models. With investigation and exploration of various methods experimented in the field of Aspect Level Sentiment Analysis, we learned the limitations and gaps in the existing research work. A major contribution of thesis is a topic modelling approach fine-tuned for the application of aspect term extraction which guarantees an almost unsupervised approach. Our initial experiments based on rule based and semantic pruning based approaches helped to ascertain the strength of these models, even though they couldn t be effective as standalone models. This progressed towards the thought of combining topic modelling approaches with prior experiments. The proposed model is an enhanced Guided Latent Dirichlet Allocation approach for aspect term extraction where the topic modelling algorithm is guided by minimal seed words corresponding to aspect categories which facilitates easier convergence. The strength of word embeddings based semantic similarity...