Lingmotif-lex: a Wide-coverage, State-of-the-art Lexicon for Sentiment Analysis

Lingmotif-lex: a Wide-coverage, State-of-the-art Lexicon for Sentiment Analysis Antonio Moreno-Ortiz, Chantal P´erez-Hern´andez University of M´alaga Spain {amo,mph}@uma.es Abstract We present Lingmotif-lex, a new, wide-coverage, domain-neutral lexicon for sentiment analysis in English. We describe the creation process of this resource, its assumptions, format, and valence system. Unlike most sentiment lexicons currently available, Lingmotif-lex places strong emphasis on multi-word expressions, and has been manually curated to be as accurate, unambiguous, and comprehensive as possible. Also unlike existing available resources, Lingmotif-lex comprises a comprehensive set of contextual valence shifters (CVS) that account for valence modification by context. Formal evaluation is provided by testing it on two publicly available sentiment analysis datasets, and comparing it with other English sentiment lexicons available, which we adapted to make this com- parison as fair as possible. We show how Lingmotif-lex achieves significantly better performance than these lexicons across both datasets. Keywords: sentiment analysis, opinion mining, lexicons, language resources 1. Introduction Sentiment Analysis has received increased attention in the last decade as a subtask of Natural Language Processing. The task can be roughly summarized as the classification of entire documents or parts of it (sentences or text seg- ments that denote a certain aspect about a certain entity). Most efforts focus on the classification of domain-specific documents, usually short texts such as product reviews and tweets, although other more sophisticated tasks are be- coming common, such as emotion and intensity detection. Traditionally, corpus-based and lexicon-based approaches have been distinguished in terms of the general system ar- chitecture. Strictly speaking, the former use a training cor- pus to extract textual cues found in each of the tagged docu- ments, whereas the latter employ a sentiment lexicon where sentiment-carrying words are stored. In practice, however, NLP practitioners combine method- ologies from both approaches, for example, (Riloff et al., 2006). Generally speaking, lexicon-based approaches are preferred for sentence-level classification (Andreevskaia and Bergler, 2007), whereas corpus-based, statistical ap- proaches are preferred for document-level classification. The use of sentiment dictionaries is a widespread method- ology, since it makes sense that the presence of certain sentiment-carrying words determine the polarity of the text in which they appear. WordNet (Fellbaum, 1998) has been a recurrent source of lexical information (Kim and Hovy, 2004; Hu and Liu, 2004; Andreevskaia and Bergler, 2006) either directly as a source of lexical information or for sen- timent lexicon construction. Other common lexicons used in English sentiment analysis research include The Gen- eral Inquirer (Stone and Hunt, 1963), MPQA (Wilson et al., 2005), and Bing Liu’s Opinion Lexicon (Hu and Liu, 2004). Yet other researchers have used a combination of existing lexicons or created their own (Hatzivassiloglou and McKe- own, 1997; Turney, 2002). The use of lexicons has some- times been straightforward, where the mere presence of a sentiment word determines a given polarity. However, negation and intensification can alter the valence or polar- ity of that word.1 Modification of sentiment in context has also been widely recognized and dealt with by some re- searchers (Kennedy and Inkpen, 2006; Polanyi …

Content truncated to 500 words