产品特征及观点的识别是细粒度情感分析的重要任务.但是,现有识别算法对中文语境下不同评论领域的适应性尚无定论,算法的鲁棒性也不理想,难以实现跨领域的算法移植.为此,选取词频统计方法、规则匹配、关联规则挖掘、具有句法格式的关联规则、CRF和SVM等6种代表性的识别算法,结合中文在线评论的语言特点,对上述算法引入到中文评论的文本分析中,根据准确率、召回率和F值指标,分析比较统计方法和机器学习方法在产品特征及观点识别上的性能.选择数码相机评论、化妆品评论、书评、酒店评论、影评、手机评论和餐厅评论7类语料3646条评论,分别采用6种算法进行产品特征和观点的抽取.实验表明,不同领域下的特征抽取难度是存在差异的;不同算法适应于不同领域;评论的文本长度对识别准确率和召回率有显著影响;另外,总体上机器学习的算法性能显著高于统计学方法.%Extraction feature and opinion is the basis of fine-grained sentiment analysis.Prior algorithms fail to be applicable for different areas,so the problem of robustness and migration for different fields are of concern to these algorithms.A couple of algorithms for feature mining have been proposed by antecedent researchers.Generally,there are two common techniques used by feature extraction:statistics based methods and machine learning based methods.However,no final conclusion has yet been drawn on the matter of feature extraction.Robustness and portability are still the main issues for current algorithms.One of the reasons is lack of systematic comparison on a unified corpus.In addition,past extraction algorithms are mostly implemented in the context of English,while lacking enough attentions on Chinese online reviews.Due to the syntactic differences between languages,the English-based algorithms cannot be directly applicable in the Chinese context.We thus choose six widely-used extraction algorithms and compare the performance between the statistical methods and machine learning methods for feature-opinion mining in Chinese context.The selected algorithms include Frequency-based opinion mining,Rule-based opinion mining,Association rule-based opinion mining,Association rule-based opinion mining plus linguistic,CRFs-based opinion mining and SVM-based opinion mining.We collect 3146 reviews as experimental corpus from 7 different fields:digital camera reviews,cosmetics reviews,book reviews,hotel reviews,critics,cell phone reviews and restaurant reviews.Finally,these corpuses are employed respectively by the six algorithms above to extract features for testing extraction performances.Experiment obtains the following conclusions:(1) It can achieve the best performance for frequency-based opinion mining when the threshold is set to 0.5％,which is quite different from English context (1％);(2) there is no algorithm which can dominate in all corpuses.Any algorithm can achieve good performance in limited areas;(3) machine learning algorithms generally outperform statistical approaches.In some corpus (e.g.mobile phone reviews),however,statistical methods can perform better,thus guiding us to select an appropriate algorithm according to the corpus in the follow-up research and application;(4) the length of reviews can affect the performance of mining algorithms.A longer text will lead a poorer accuracy,and vice versa;(5) due to syntactic difference between languages,both the association rule-based opinion mining and the association rule-based opinion mining plus linguistic perform poorly in Chinese context,unlike their excellence in English context.It also implies the complexity of Chinese natural language processing;(6) for the same algorithm,experimental results can be better in dealing with service domains (e.g.restaurants,hotels),but much poor in dealing with arts and entertainment area (e.g.film,book).It indicates the differences between domains in problem solving of feature extraction.
展开▼
