Semantic Dependent Word Pairs Generative Model for Fine-Grained Product Feature Mining

Abstract

In the field of opinion mining, extraction of fine-grained product feature is a challenging problem. Noun is the most important features to represent product features. Generative model such as the latent Dirichlet allocation (LDA) has been used for detecting keyword clusters in document corpus. As adjectives often dominate review corpus, they are often excluded from the vocabulary in such generative model for opinion sentiment analysis. On the other hand, adjectives provide useful context for noun features as they are often semantically related to the nouns. To take advantage of such semantic relations, dependency tree is constructed to extract pairs of noun and adjective with semantic dependency relation. We propose a semantic dependent word pairs generative model for pairs of noun and adjective for each sentence. Product features and their corresponding adjectives are simultaneously clustered into distinct groups which enable improved accuracy of product features as well as providing clustered adjectives. Experimental results demonstrated the advantage of our models with lower perplexity, average cluster entropies, compared to baseline models based on LDA. Highly semantic cohesive, descriptive and discriminative fine-grained product features are obtained automatically.