Semantic Orientation Analyses On Web Opinion

With the widespread of Internet and developing of Internet Technology, human are closely related to Internet. More and more people shopping online and express their opinions online, and people can surf the Internet anywhere and anytime as a result of developing of mobile Internet. This cause the amount of data online increasing fast.Opinion mining aims at processing large amount of data and extract the valuable information for people. There are three levels of mining, document level, sentence level and word level. Each level adapts to different situation.This paper concentrates on document level opinion mining, also called semantic orientation analysis, which aims at judging whether a document is positive, negative or objective. With considering of the popularity of online shopping, the importance of mobile devices in people’s daily life, this paper decide to analysis the comments on mobile phones.This paper analyses the semantic orientation using the method based on sentiment dictionary. Build basic positive dictionary by combining HowNet Positive Evaluating Words, HowNet Positive Semantic Words and the positive words in Semantic Ontology. Build basic negative dictionary by combining HonNet Negative Evaluating Words, How Net Negative Semantic Words and the negative words in Semantic Ontology. In order to make our method more applicable to mobile phone domain, we have done much corpus research to extract the words that are domain related and extend our sentiment dictionaries with the words extracted. This makes our sentiment dictionaries more applicable in mobile phone domain.Assistant dictionary set includes five dictionaries: stop-dictionary, privative-dictionary, adverb-dictionary, relate-dictionary and dynamic-dictionary. Use stop-dictionary to filter the words that doesn’t have real meaning. Use privative-dictionary to handle the privative words. Recording the number of privative words when matching. If the number is odd, this means that the privative words express negative and changes the orientation of the sentiment word. If the number is even, this means that the privative words express positive and double the orientation of the sentiment word. Use adverb-dictionary and relate-dictionary to handle the adverb and relate words. Use dynamic-dictionary to handle the changing orientation of some sentiment words. When matches a dynamic subject, change the orientation of sentiment word to the opposite.When processing the Orientation analysis, match sentiment words after pre-processing. If succeed, traversal the sentence and match adverbs according to adverb-dictionary. Double the orientation when an adverb matched. Traversal sentence and match privative words. Change the orientation according to the number of privative words. If a relate word is matched, multiple the orientation by a factor. If a dynamic word is matched, change the orientation of sentiment word to the opposite. In positive analysis, this paper gets a precision of 71.3% and a recall of 76.9% in mobile phone domain. And in negative analysis, this paper gets a precision of 71.3% and a recall of 76.9%.This paper also presents a supervised method to find out new words. A threshold value is preset. If a word doesn’t match any word in all those dictionaries above, record the number that the word occurred, the pre-piece and the after-piece. While the number of a word is greater than the threshold value, present the word and the pre-piece, after-piece to human to judge whether there is a new word. Plus all orientation value of sentiment words, get the orientation value of sentence. Plus all the sentences’ orientation value, get the orientation value of the whole document.