SenSALDO: Creating a Sentiment Lexicon for Swedish

SenSALDO: Creating a Sentiment Lexicon for Swedish Jacobo Rouces, Nina Tahmasebi, Lars Borin, Stian Rødven Eide Språkbanken, University of Gothenburg {jacobo.rouces, nina.tahmasebi, lars.borin, stian.rodven.eide}@gu.se Abstract The natural language processing subfield known as sentiment analysis or opinion mining has seen an explosive expansion over the last decade or so, and sentiment analysis has become a standard item in the NLP toolbox. Still, many theoretical and methodological questions remain unanswered and resource gaps unfilled. Most work on automated sentiment analysis has been done on English and a few other languages; for most written languages of the world, this tool is not available. This paper describes the development of an extensive sentiment lexicon for written (standard) Swedish. We investigate different methods for developing a sentiment lexicon for Swedish. We use an existing gold standard dataset for training and testing. For each word sense from the SALDO Swedish lexicon, we assign a real value sentiment score in the range [-1,1] and produce a sentiment label. We implement and evaluate three methods: a graph-based method that iterates over the SALDO structure, a method based on random paths over the SALDO structure and a corpus-driven method based on word embeddings. The resulting sense-disambiguated sentiment lexicon (SenSALDO) is an open source resource and freely available from Språkbanken, The Swedish Language Bank at the University of Gothenburg. Keywords: sentiment analysis, Swedish, lexicon, lexical resource 1. Introduction The natural language processing (NLP) subfield known as sentiment analysis or opinion mining has seen an explosive expansion over the last decade or so. Since the publication of the comprehensive overview of the field by Pang and Lee (2008), we have seen hundreds of papers as well as dedicated workshops on this topic in NLP conferences. Even though sentiment analysis has become a standard im- plement in the NLP toolbox, many theoretical and method- ological questions remain unanswered and resource gaps need to be filled. Most work on automated sentiment anal- ysis has been done on English and a few other languages; for most of the written languages of the world,1 this tool is not available. This paper describes the development of an extensive sentiment lexicon for written (standard) Swedish, an essential component in sentiment analysis. The theoretical and methodological issues that arise in con- nection with sentiment analysis of texts lie partly in the intersection of the linguistic subfields of pragmatics and lexical semantics. Depending on your view of the scope of these subdisciplines, you may end up with very differ- ent thoughts about the prior-polarity – i.e., lexical-semantic – and contextual – i.e., pragmatic – elements of sentiment information, and how these are combined in concrete text analysis. An added complication is that sentiments and emotions are central objects of study also in other fields, notably psychology. In practice this means that we find many different propos- als in the literature, for how prior sentiment polarity should be represented in the lexicon, which kinds of lexical enti- ties should be attached (lemmas, lexemes or word senses), and how contextual information …

Content truncated to 500 words