Finer Grained Opinion Analysis On Product Reviews

In recent years, with the blooms of Internet, there comes out a lot of subjective texts.Especially for the product evaluation texts, it’s encouraged to share comments of productevaluation early by the electricity supplier. As an important subtask of sentiment analysis,fine-grained emotion analysis can analyze the specific emotional elements in subjectivetexts, which has gained more and more attention by researchers.In previous researchs of grained emotional analysis, some approachs use the tem-plates to extract the fine-grained elements, however, which has a poor flexibility and lowrecall for extraction. Many methods are fail to classify the polarity of the pair of theproduct attribute and the key word of opinion expression.To meet the need for finer-grain sentiment analysis, we proposed a strategy of thecorpus annotation system which organizes the product attributes in the domain ontologystyle. Then manually annotate1000reviews crawled from the digital cameras web sitesto be a set of open corpus, which notes the appearance of the pair of product attribute andthe key word of opinion expression that provide data support for fine-grained sentimentanalyse. In this article, a semi-supervised learning approach is presented to extract theinstances of the product attributes. In practical corpus, especially for the corpus of specificdomain, the instances of product attribute are always explained in a out dictionary way,for which general segmenter performs poor. This approach can mining the out-dictionarywords and provide the feature support to the further domain-specific analyse. We alsepropose to apply tree-structured CRFs model to extract the sentiment elements. Insteadof the word sequence, the use of dependency features is helpful to describe the directrestrictions which are insensitive to the word distance in the sentence. The experimentalresults on COAE2011dataset and CUHK Opinmine dataset show that the proposed tree-structured CRFs achieves better performance comparing with CRFs without edge featuresand CRFs with linear features.