While machine learning are widely used in sentiment analysis, there are also many sentiment analysis systems adopting unsupervised learning methods. Where sentiment pairing words and phrases are collected and then searched for during analysis. And this way, we can come up with a certain sentiment index. There are two types of Lexicons. First one is sentiment dictionary. Popular dictionaries include Harvard General Inquirer, LIWC, MPQA Subjectivity Cues Lexicon, SentiWordNet. Second one is Corpus-based sentiment lexicon. Which, building specialized sentiment dictionaries, from the corpus of text. With lexicons, either clustering or scoring functions or algorithms are used to determine sentiment of the input text. The Harvard General Inquirer is a lexicon attaching syntactic, semantic and pragmatic information to part of speech type words the home page for general inquiry is www.wjh.harvard.edu/-inquirer. General inquire is basically a mapping tool It maps each text file with counts on dictionary supplied categories. The content distributed version combines various resources including positive and Negative. The Positive has 1915 words and negative has 2291 words. The Harvard IV-4 dictionary content-analysis categories, that also, it includes Lasswell dictionary content-analysis categories. And the last category is five categories based on social cognition work of Sentiment and Feeders which makes for 182 categories in all. In the category is a list of words, and word sentences. The dictionary is freely downloadable, for research purpose. LIWC stands for Linguistic Inquiry, and word count, and W LIWC is a popular database consisting of lots of categorized, regular expressions. The home page of LIWC is www.liwc.net. The LIWC2007 dictionary contains 4500 words and word stems. Each is filed into one or more sub dictionaries and sub dictionaries represent one of the 55 word categories through which LIWC comprise a text. For example the word cried is part of five word categories, sadness, negative emotion, overall affect, verb and past tense verb. Hence, it is found in the text each of these five sub dictionary scale scores will be incremented. It costs around $90 and its classifications are highly co-related with those of Harvard General Inquirer Dictionary. MPQA stands for Multi-Perspective Question Answering. The homepage of MPQA is www.cs.pitt.edu/mpqa/subj_lexicon.html. MPQA has several different sentiment dictionaries including MPQA opinion corpus, subjectivity lexicon, subjectivity sense annotation, arguing Lexicon and the last one is, plus/minus effect lexicon. Let me explain each of those briefly. The MPQA Opinion Corpus contains news articles from wide variety of news sources manually on a data for opinions. It consists of around 10,000 sentences from the WordPress, annotated for subjective expressions. The subjectivity lexicon, contains a list of subjectivity clues. The lexicon includes patterns that present arguing 17 files that represents a type or category of arguing. SentiWordNet attach positive and negative real value sentiment scores to WordNet WordNet's synsets and it's hierarchies. So let me talk briefly about WordNet. WordNet is a general anthology developed by Miller and his colleagues at Princeton University. SentiWordNet assigns to each synset of WordNet three sentiment scores, positivity, negativity, and objectivity. Current version of SentiWordNet is 3.0. I'll talk more about SentiWordNet in next session. Dictionary based approaches, typically use words like synset and hierarchies to acquire opinion words. Many sentiment lexica including ones that I just covered in this session can be found on the web. They often have thousands of terms and they are quite useful for many different tasks. There are several weaknesses of dictionary-based approaches. Lexicon-based sentiment analysis systems are hard to develop. High quality lexica resources are key for good performance. A cognition process is hard. Usually does not give domain or context dependent meaning. Those are the sum of of dictionary based approaches. To overcome those with of dictionary based approaches, corpus based approach or proposed. In general, corpus-based approach relies on syntactic or co-occurrence patterns in large corpuses. It also requires a large corpus to get good coverage. Surprisingly It can find too many independent orientation such as positive, negative, neutral. Turney in 2002 and Yu and Hazivassiloglou, suggest methods to determine the polarity of document using word similarity. They are, both of them, assigned opinion orientations which are polarities to words or phrases. To preserve sentiment consistency a typical approach is to use conventions on connectives to identify opinion words. For example, conjunction is used. Since, conjoined adjectives, usually have the same orientation. Let's take the following example sentence. This car is beautiful, and spacious. Here, beautiful and spacious conjunctions. Learning-based model is used to determine if two conjoined adjectives are of the same or different orientations. Alternatively, clustering techniques can be used to produce two sets of words, positive and negative.
