| Comparison is one of the most common expressions, which is widely used by people toexpress their emotional tendency on different things. With the rapid development of computertechnology, there are a large number of Uyghur social and business application platformappear in the Internet, and the Uyghur texts of the platform contain many comparativesentences. As one of the most important parts of fine-grained opinion mining, identifyingcomparative sentences and extracting comparative relations have important practicalsignificance. Comparative opinion mining is related to several technologies, such as textclassification, information retrieval and information extraction. The information mined by uswill offer valuable reference for some potential buyers or investors. However, there is noresearch on Uyghur comparative sentences, and for the existing methods on miningcomparative opinions, they have quite low recall rate.This paper sets comparative texts as the object and studies the comparative opinionmining method, and it includes the following two aspects:(1) Identification of comparative sentences. We put forward a two-level model to identifysentences. The first level is to identify sentences based on our comparative words set; and thesecond level uses Bidirectional CSR Mining algorithm (Bi-CSR) that we propose in this paperto mine comparative patterns, and take them as feature for Support Vector Machine (SVM) toclassify into either comparative or not.(2) Extraction of comparative type and relation. For the type identification, we settype-indicating words in an order, and classify sentences based on these ordered words.Relation extraction is to get the comparative subject (Subject), comparative object (Object),feature (Feature) and comparative result (Result) from a comparative sentence. In order to getthese elements, we propose different methods for different kind of elements. Firstly, we treatSubject and Object as the topic of an opinion, and identify them through Latent Dirichlet Allocation (LDA) model. Then we use similarity calculation and a mapping table of Featureand Result to classify topics we get from LDA into Subject, Object and Feature. Lastly, wetake full use of the features of Uyghur comparative sentences to generate rules for Resultextraction. After we get all the above elements, we express comparative relation in the form of5-tuple <Subject, Object, Feature, Result, Type>.We do experiment according to the above idea, and through the experiment results weget the conclusion that our method can solve the problem of Uyghur comparative sentenceidentification and relation extraction. Our method gets good performance in both precisionand recall rates. |
