Generating domain ontology from Chinese customer reviews to analysis fine-gained product quality risk

With the rapid development of E-commerce in China, quality of the products on online shopping platforms has caused wide concern. Customer reviews, which commented by people who bought the very product, now have been one of the most important resources for analyzing product’s quality risk. We can get fine-gained, aspect-oriented risk information of a product by mining its reviews. Unfortunately, people tend to write reviews with casual grammar or just omit parts of components of a sentence. Both these features will cause negative impacts when parsing the raw customer reviews directly. Thus a knowledge base which is built totally beyond the reviews could be used to analyze it despite the drawbacks above. In this paper, we generate a domain ontology from raw text in the online encyclopedia. It can be viewed as a graph whose nodes represent domain concepts and edges represent the relations between these concepts. In our work, we integrate syntactic tree structure in linear-chain CRFs for recognizing domain concepts and train SVMs and MaxEnt models on elaborate features for clarifying three types of relationship, namely “Attribute-of”, “Part-of” and “Instance-of”. Once the ontology has been built, product properties with potential risk will be extracted by our matching method. Experiment show that our approach achieves 64.4% precision and 82.4% recall on risky property extraction task.