Domain-Specific Sentiment Lexicons Induced from Labeled Documents

Proceedings of the 28th International Conference on Computational Linguistics, pages 6576–6587 Barcelona, Spain (Online), December 8-13, 2020 6576 Domain-Specific Sentiment Lexicons Induced from Labeled Documents SM Mazharul Islam University of Texas at Arlington sxi7321@mavs.uta.edu Xin Dong Rutgers University–New Brunswick xin.dongx@rutgers.edu Gerard de Melo Hasso Plattner Institute, University of Potsdam http://sentimentanalysis.org gdm@demelo.org Abstract Sentiment analysis is an area of substantial relevance both in industry and in academia, includ- ing for instance in social studies. Although supervised learning algorithms have advanced con- siderably in recent years, in many settings it remains more practical to apply an unsupervised technique. The latter are oftentimes based on sentiment lexicons. However, existing sentiment lexicons reflect an abstract notion of polarity and do not do justice to the substantial differences of word polarities between different domains. In this work, we draw on a collection of domain- specific data to induce a set of 24 domain-specific sentiment lexicons. We rely on initial linear models to induce initial word intensity scores, and then train new deep models based on word vector representations to overcome the scarcity of the original seed data. Our analysis shows sub- stantial differences between domains, which make domain-specific sentiment lexicons a promis- ing form of lexical resource in downstream tasks, and the predicted lexicons indeed perform effectively on tasks such as review classification and cross-lingual word sentiment prediction. 1 Introduction Sentiment analysis is among the most prominent forms of natural language processing, with applica- tions such as social media analytics (Rosenthal et al., 2017; Wang et al., 2019; Shoeb et al., 2019), marketing and customer support (Gamon, 2004), as well as recommendation (Yang et al., 2013). Apart from machine learning-driven systems (Pang et al., 2002; Socher et al., 2013; Kalchbrenner et al., 2014, inter alia), which require supervision using labeled training data, there are also lexical resource-driven systems that exploit sentiment lexicons and can be run out-of-the-box without the need for any labeled training data. Well-known sentiment lexicons include the Hu and Liu (2004) Opinion Lexicon, Senti- WordNet (Baccianella et al., 2010), LIWC (Pennebaker et al., 2001), and VADER (Hutto and Gilbert, 2014). There are numerous techniques for lexicon-driven sentiment analysis (Taboada et al., 2011), Sen- tiStrength (Thelwall et al., 2010) being an example of a more modern lexicon-driven sentiment analysis system. Sentiment lexicons can also be used to bootstrap domain-specific supervised sentiment analysis models (Mudinas et al., 2018). A sentiment lexicon is a resource that, for a given word (form) w, provides an annotation label lw describing its overall sentiment polarity. Some lexicons merely provide labels in {positive, negative} or {positive, neutral, negative}. Others offer more informative intensity scores to account for the fact that some words are more negative or positive than others. For example, an emphatic word such as spectacular is generally considered stronger than a simple good (de Melo and Bansal, 2013). Such scores could be in the range [−1, 1], with −1 denoting the most negative sentiment polarity, whereas +1 is the most positive score. In this paper, we consider …

Content truncated to 500 words