Sentiment Analysis Of Web Financial Reviews Based On Syntax And Semantic Mining

With the rise of Web 2.0,the Internet has developed from a top-to-down network system controlled by a small number of resource controllers,into a down-to-top network system which is spontaneously and unanimously dominated by the majority of users.Meanwhile,with the rapid development of Internet technology,the population of netizens has continually increased,as has convenient access to the Internet.Thus,the Internet has become an important place for people to leave comments.In the current network system,combined with huge numbers of netizens with convenient access and the willingness of people to share via and trust the Internet,numerous subjective comments have emerged,including comments regarding news and politics,online commodities,public figures,economic developments,etc.The sentiment analysis of comment texts can be widely applied to numerous fields such as information retrieval,answering systems,commercial intelligence,economic forecast,public opinion,etc.The sentiment analysis of comment texts has become a research hotspot in the fields of data mining,computer linguistisc,and artificial intelligence.Meanwhile,due to its significance in applications,it has become a focus of the industrial realm.Large amounts of text data describing the financial sector has also emerged on the Web,i.e.,Web financial reviews.Sentiment mining oriented to Web financial reviews can be an important index for the financial early warning(FEW)of listed enterprises,and it can also be used to predict stock price trends.Currently,sentiment analysis for reviews primarily focuses on the field of commodity remark,while studies on sentiment analysis for Web financial reviews are still in the initial stages.Different from sentiment analysis for commodity reviews,the difficulties of sentiment analysis for financial reviews lie in the following aspects:(1)The number of opinion targets is large and the composition of the objects is complex.The opinion targets in commodity reviews are primarily nouns or noun phrases;while in financial reviews,the opinion targets may also be clauses,such as subject-verb phrases,verb-object phrases,etc.(2)The frequency difference of objects is large.In commodity reviews,the distribution of opinion targets is relatively even.In financial reviews,the numbers of comments in the data source may significantly differ,such as the analysis of financial indicators in financial statements,or the review of non-financial indicators such as employee satisfaction.This makes the occurrence frequency of opinion targets very different.(3)The syntactic constituents of opinion words in sentences are very flexible.In commodity reviews,opinion words are primarily adjectives.In financial reviews,the part-of-speech(POS)to which opinion words belong is more diverse,including adjectives,verbs,and nouns.Verbs are expecially frequent opinion words.(4)Ambiguous opinion targets are more common.Objects in financial reviews span a wide range and appear in complicated composition forms,but the Chinese language requires them to be expressed in concise terms.Therefore,abbreviations or reference forms of words are often used.These kinds of objects are called virtual opinion targets.(5)Implicit opinion targets are more common.Due to the language characters of the Chinese language,subjects or objects are often omitted in sentences with semantic context cues,or the opinion targets are implicitly contained within hints provided by particular words.(6)There are more singular opinion targets.In commodity reviews,the opinion targets are generally the features or properties of products,which are primarily represented by nouns or noun phrases without sentimental polarities.Due to the complexity of opinion targets in financial reviews,they may contain sentimental polarities such as the noun phrase opinion targets "increase","decrease",etc.Clause opinion targets may also demonstrate sentimental polarities due to predicates.(7)It is more common to express the sentimental degree by numbers.The degree modifier of the emotional words is more adverbs in product reviews,but it is more common to express the sentimental degree by numbers except adverbs in financial reviews.(8)The documents and sentences of financial reviews are longer.Commodity reviews typically contain only one sentence which contains a review of each property of the commodity.However,the description of financial reviews is more professional,with more complex and longer sentences.Fine-grained sentiment analysis of Web financial reviews is thus a huge and complicated project.Focusing on the above characters of Web financial reviews,the following aspects of study are conducted:(1)The generalization of nine syntactic constituents that affect emotional propensity and eight pairs of dependencies.It is summarized that opinion words can fall into nine types of syntactic constituents in a sentence,according to their part-of-speech and the syntactic constituents in sentences.Additionally,the effects of 24 types of dependency relation on sentence sentiment computing are analyzed.Eight types of dependency relation which may affect the sentiment tendency of sentences are identified.(2)Extracting target-opinion pairs and extending the opinion targets.A target-opinion pair is denoted as<Opinion Target,opinion word>,and is the combination of an opinion word and corresponding opinion targets.This study mines the opinion targets corresponding to opinion words from the perspective of shallow semantics and syntactic analysis.Due to the existence of ambiguous opinion targets,ambiguous opinion targets are determined and replaced in this study by using domain knowledge such as financial dictionary,financial indicators,non-financial indicators,and basic enterprise conditions.There are different reasons for the occurrence of default and implicit opinion targets.Based on this,three corresponding recognition methods for implicit opinion targets have been determined for three possible situations.(3)Sentimental polarity unit extraction based on syntactic analysis and recognition of singular opinion targets.A sentimental polarity unit is referred to as<Opinion Targets,N,opinion word>.The sentimental polarity of opinion targets is not only affected by the original polarities of opinion words,but also by the modification singularity of negative words and the dynamic polarities of singular opinion targets.In this study,because the parallel structure of opinion targets represents the consistence of the polarities of two opinion targets and because the sentimental polarities of the two clauses separated by a transitional word in a sentence are opposite,the singular opinion targets are identified.(4)Sentiment calculation for financial reviews based on syntactic analysis.The sentimental tendency includes two aspects:sentiment polarity and sentiment strength.The factors that affect sentiment polarity include opinion words,negative modification and opinion targets.The factors that affect sentiment strength include degree adverbs,the edit distances between negative words and opinion words and the edit distances between degree adverbs and opinion words.By focusing on the six factors affecting sentiment tendency,a fine-grained sentiment tendency computing model for Web financial reviews is constructed.(5)We conduct comprehensive experiments on Chinese financial reviews,and experimental results show that the proposed methods are effective.The novelty of this study lies in the following aspects:(1)Based on shallow semantic and syntactic parsing,we propose a new method for extracting target-opinion pairs from Chinese financial reviews.(2)Based on domain knowledge and context semantic knowledge,we propose a new method to solve ambiguous opinion target and implicit opinion target.(3)Based on sentence grammar analysis,we propose a new method to identify the emothional polarity of singular opinion target.(4)Based on the part-of-speech tagging and dependency parsing,we build the fine-grained sentiment computing model for Web financial reviews.