基于情感本体的在线评论情感极性及强度分析:以手机为例

Sentimental polarity and strength of online cellphone reviews based on sentiment ontology

对在线评论进行情感分析,有利于消费者制定购买决策,也有利于商家确定营销策略。已有对在线评论的研究,多单独采用语义方法或统计方法进行单粒度的情感分类。语义方法借助已有情感词典,忽略了上下文语境,影响分类准确率;统计方法需要对大量语料进行人工标注,影响分类效率;只针对单粒度语料分类,不能同时获得用户对产品整体和细节特征的观点,影响分类结果的应用性。针对以上问题,本文提出一种基于情感本体的在线评论情感极性及强度分析方法。首先,根据已有在线声誉系统和中文网络词汇特点,结合语义方法和统计方法,通过特征观点对的抽取和观点词情感的判断,构建情感本体。情感本体的构建既考虑了上下文语境,又不需要预先的语料标注;其次,通过情感本体的应用,实现了在线评论的评论整体和具体特征的情感极性及强度分析。通过对手机评论的实验,结果显示,通过应用该情感本体,可以有效地给出用户对产品整体和属性细节的满意或不满意的态度。

With the development of Intemet and E-commence, an increasing number of people have submitted or retrieved online reviews about products via avariety of web-based channels. Online reviews can facilitate consumers＇ purchasing decisions and merchants＇ sales decisions because online reviews reflectusers＇ opinions on certain products. However, it is impossible to collect and summarize users＇ opinions by hand because of the explosion of unstructured onlinereviews. Therefore, the demand for sentiment classification rises in response to the requirement of automatically retrieving valuable information from onlinereviews. There are two kinds of approaches for sentiment classification： semantic approach and statistical approach. The existing researches only performsentiment classification at single levels of granularity based on semantic approach or statistical approach. The semantic approach performs context-fleesentiment classification based on the existing sentiment lexicon, while the statistical approach performs context-sensitive sentiment classification based on alarge number of manually annotated training reviews. Therefore, it is necessary to combine statistical approach and semantic approach to improve theeffectiveness of sentiment classification. Traditional sentiment classification was performed at the coarse-grained level. A growing number of studies haveexamined sentiment classification at the more fine-grained phrase level in recent years. To identify consumers＇opinions on both coarse-grained level andfine-grained phrase level, and to improve the applications of sentiment classification, tackling the sentiment classification problem at varying levels ofgranularity is necessary. To tackle these problems, we develop sentiment ontology to facilitate sentiment classification at varying levels of granularity and assess the sentimentalpolarity and strength of online reviews. Then, the effectiveness of the sentiment ontology is rigorously evaluated based on real online reviews retrieved from apopular E-Commerce website. Firstly, according to the characteristic of existing online reputation system and Chinese online reviews, we apply the combinationof statistical approach and semantic approach to automatically construct sentiment ontology. The sentiment ontology is mainly focused on extractingFeature-Opinion Pair and identifying sentiment of opinions. After that, sentiment ontology is used to improve the effectiveness of sentiment classification onboth the coarse-grained level and fine-grained phrase level. At last, we conduct experiments to quantitatively evaluate the effectiveness of the sentimentontology based on real cellphones online reviews. The experimental results of sentimet ontology construction show that sentiment values of opinion words are different in different Feature-Opinion Pairs.Negative words not only change the polarity of opinion words but also weaken sentiment intensity of the words. Degree adverbs strength the sentimentintensity of opinion words, but the degree of strengthening declines with the increasing intensity. The experimental results of sentiment classification show thatsentiment ontology is an effective solution for sentiment classification on both the coarse-grained and fine-grained phrase levels. This paper utilizes sentimentontology as a feature dimension reduction method. As a result, the accuracy of classification is higher than traditional feature dimension reduction method. Theimproved accuracy proves that sentiment ontology is effective for coarse-grained level sentiment classification. The sorting of sentiment intensity valuescalculated by sentiment ontology is the same as the sorted values with manual analysis. Sentiment intensity values prove that sentiment ontology is effective forfine-grained phrase level sentiment classification.