Research On The Default Identification And Recovery Of Evaluated Element

With the increasingly openness of the social networking platform, increasing comment text data presents the characteristics of semi-structure, colloquial, irregular etc., then the phenomenon of default becomes more and more popular. In opinion sentence, the default item of evaluated object and attribute can make the expression more concise, meanwhile it brings much uncertainty problems for the NLP technology based opinion mining. With the increasingly deepening of related study of grained opinion mining, the problem of default of evaluated factors has attracted the attention of researchers. In order to solve the default of evaluated object and attribute, around the identification and recovery for the default of evaluated object and attribute, this paper carries out two tasks, and systematically puts forward a scheme for the identification and recovery of default. The major works of this thesis include:(1) The identification method for default item of evaluated object and attributeFrom the perspective of syntactic constituents, the rule set of default item identification is constructed to obtain the candidate set of recognized default item. On the basis, we regard the identification of default item as a binary classification problem, and import the lexical and dependency parsing features. We employ the decision tree C4.5 algorithm to train classification model which is used to judge the recognized default item on the testing data. Experimental result shows that the accuracy of the integrating of two features sets of lexical and dependency parsing is 66.3%, which is superior to each single feature.(2) The determining for the type of default item of evaluated object and attributeIn order to accurately realize the recovery the default item of evaluated object and attribute, we need the determining algorithm which can provide guidance cues for the recovery of default item. According to the distribution of default item of evaluated object and attribute in corpus, We put forward the method of the rule matching and dynamic constructing attribute-indicator (A-I), respectively. Through the experiments on the car reviews and mobile phone micro-blog show that the recall of car and mobile phone in the determining for the type of default item of evaluated object are respectively 92.1% and 67.8%, the recall of car and mobile phone in the determining for the type of default item of evaluated attribute are respectively 91.8% and 78.0%, it shows that the method of this thesis is suitable for the car review data, furthermore, comparing the forum reviews, micro-blog is more irregular.(3) The recovery for default item of evaluated object and attributeAiming at the recovery for default item of evaluated object and attribute, On the basis of determining the type of default item, we design three Strategies for recovering default item, and adopt the nearest neighbor matching method to obtain the default evaluated object. For the recovering default item of evaluated attribute, we use the method of A-I to acquire the default evaluated attribute. Through the experiments of recovering the default evaluated object on the car reviews and mobile phone micro-blog show that the accuracy of car and mobile phone are respectively 92.1% and 67.8%, the accuracy of recovering default evaluated attribute are respectively 92.1% and 67.8%, it shows that the recovering the default evaluated object is better on the situation of less evaluated object. The effect of the recovery for default evaluated attribute of car is superior to mobile phone, it shows that the greater corpus, the more acquiring the information of A-I, and the effect of recovering the default evaluated attribute is better.