Incorporating Dependency Analysis Into Topic Modeling For Fine-grained Opinion Mining

With the considerable growth of web2.0, reviews have become valuable sources for mining customers’opinions. Expressing opinions through online reviews has a great deal of flexibility and complexity. However, most of the traditional opinion mining methods are coarse-grained and cannot understand the natural language. To solve these problems, this paper presents two fine-grained opinion mining models based on syntactic analysis and topic modeling. The models study the automatic extraction of product and service aspects, as well as the sentiment words from reviews.First, an unsupervised approach is presented for extracting the Appraisal Expression Patterns (AEP) from the reviews. AEP represent the semantic relationships between aspects and sentiment words. They are high-level language-dependent types of semantic information and have excellent domain adaptability. Second, an AEP-based Latent Dirichlet Allocation model (DLDA) is proposed. DLDA is a sentence-level probabilistic generative model. It assumes that every review dataset is composed of several mutually corresponding domain and sentiment topics, as well as a background word topic. The DLDA model can use AEP information to simultaneously mine accurate aspects and sentiment words. However, it uses only AEP and cannot take full advantage of the other features, thus, it has poor scalability. Finally, a Maximum entropy-based LDA model (MLDA) is proposed. MLDA is a supervised probabilistic generative model and it has a better scalability due to the introduction of maximum entropy model. However, MLDA model needs annotated training data.Experiments were conducted in the domain hotel, restaurant, MP3player and digital camera. Experiments show that, the DLDA model is better than the baselines in the aspects extraction, sentiment words identification and domain adaptation. MLDA also outperforms two classical approaches in the extraction of aspects.