Research On The Identification Approach Of Opinion Element Orientation For Chinese Comparative Sentences

With the network turning from information unidirectional provider to multi-platform,more and more users transfer from pute network information readers to network information's creators and actors.Through the kinds of network platform such as shopping websites,users create,share and consume contents,these amounts of contents and comments are provided to other users to recognize things they are not familiar with before,which affects other users' recognition.Among the rest is comparative sentences.People are getting used to taking other users' comments via network to help himself compare the goods he interested and make the decisions finally.In the era of big data,using machine learning to mining these comparative sentences can help us get the comparing data among two or more goods automatically,gain the valuable information more quickly.This thesis mainly analyzes and studies the characteristic of Chinese comparative sentences,and recognizes the comparative sentences and comparative sentences' opinion tendency.In the comparative sentences recognition task,this thesis used Chi-square detection method to select the text features,used TF-IDF to vectorize the texts,and used SVM classifier to archive the comparative sentences recognition.In the comparative sentences' opinion tendency recognition task,this thesis combined the CRF and SRL methods to complete the research.The main research description are in the following:(1)Used Chi-square detection method for “words/part of speech of this words” to select the text features,and expand some comparative words as features.(2)Based on machine learning methods(SVM)to classify the comparative sentences.(3)Used Sequence Labeling and CRF to extract the comparative sentences' comparing objects.(4)Used SRL to extract the comparative sentences' objects.(5)Built the basic emotion dictionary,area attributes dictionary and negative word dictionary,to identify the comparative sentences' opinion tendency.This thesis used kinds of machine learning's experiment for comparative sentences' training materials and testing materials provided by COAE 2012 and COAE 2013.In the comparative sentences recognition task,considering the comparative sentences' characteristic,used Chi-square detection to pick up text features,and used SVM to classify the sentences.In electronics field,this method got good performance with recall 0.843,precision 0.999 and F1 value 0.914 for comparative sentences' recognition,meanwhile got the good performance with recall 0.999,precision 0.864 and F1 value 0.927 for non-comparative sentences' recognition,both comparative sentences' recognition and non-comparative sentences' recognition got the best performance in all contestants.In car field,the performance's recall and F1 value are also better than all contestants.After adding other comparative feature words into text features creatively,the performance was further improved to recall 0.972,F1 value 0.967 for electronics field,and recall 0.941,F1 value 0.946 for car field.On macro-average and micro-average side,the recall is 0.15 points higher than the best performance in all contestants and the F1 value is 0.07 points higher than the best,which show the great effectiveness of this method.In the comparative sentence opinion element orientation's recognition task,the thesis combined the CRF and SRL methods to extract the comparative sentences' objects and opinions,and took the customized emotion dictionary to identify the opinions tendency.By using CRF method to extract the objects,the recall is 0.6 above in both electronics and car field,which is better than all contestants.And the recall performance got improved after combined the results using two different CRF templates to train and test.After continued to combine the SRL's extraction result,the recall of objects' recognition and attributes' recognition were got further improved.Although the recall and F1 value are lower than the best performance in the finial opinion element orientation assessment,the precision got the best performance,which show the feasibility of this method.