Computing with Subjectivity Lexicons

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3272–3280 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3272 Computing with Subjectivity Lexicons Caio L. M. Jeronimo, Claudio E. C. Campelo, Leandro Balby Marinho, Allan Sales Adriano Veloso, Roberta Viola UFCG - Brazil, UFMG - Brazil Campina Grande - Brazil, Belo Horizonte - Brazil {caiolibanio, allanmelo}@copin.ufcg.edu.br, {campelo, lbmarinho}@computacao.ufcg.edu.br adrianov@dcc.ufmg.br, robertaviola@ufmg.com Abstract In this paper, we introduce a new set of lexicons for expressing subjectivity in text documents written in Brazilian Portuguese. Besides the non-English idiom, in contrast to other subjectivity lexicons available, these lexicons represent different subjectivity dimensions (other than sentiment) and are more compact in number of terms. This last feature was designed intentionally to leverage the power of word embedding techniques, i.e., with the words mapped to an embedding space and the appropriate distance measures, we can easily capture semantically related words to the ones in the lexicons. Thus, we do not need to build comprehensive vocabularies and can focus on the most representative words for each lexicon dimension. We showcase the use of these lexicons in three highly non-trivial tasks: (1) Automated Essay Scoring in the Presence of Biased Ratings, (2) Subjectivity Bias in Brazilian Presidential Elections and (3) Fake News Classification Based on Text Subjectivity. All these tasks involve text documents written in Portuguese. Keywords: subjectivity, lexicon, word embedding 1. Introduction Many everyday tasks in Natural Language Processing (NLP) rely on lexicons, which are vocabularies that repre- sent different branches of knowledge. Lexicons are useful resources for identifying semantics relevant to sentiment, emotion, personality, language bias, mood, and attitude. Using lexicons in computational tasks has enabled several research works and exciting new applications to arise, such as detecting sentiment towards politicians and products, frustration of callers in call centers, stress in drivers or pi- lots, depression, media bias, biased ratings in essay scoring, and fake news classification. Choi and Wiebe (2014), for example, develop a sense-level lexicon to support opinion mining. The authors combine a graph-based method, based on WordNet relations, with a standard classifier, trained on gloss information, and show that the model can be useful to guide manual annotations to find positive/negative senses. Recasens et al. (2013) identify familiar linguistic cues related to framing and epis- temological bias, which include factive verbs, implicative, hedges, and subjective intensifiers. From these cues, the authors developed a bias lexicon for extracting features for a model to solve the problem of bias-inducing word identi- fication. The recent advances in Deep Learning have also brought relevant changes in the way lexicons are constructed and exploited. Zeng et al. (2018), for example, expand the well known LIWC (Linguistic Inquiry and Word Count) (Pen- nebaker et al., 2001) resource for the Chinese language, by the usage of Sequence-to-Sequence models and word embeddings to classify words in the lexicons. The au- thors show that their model outperforms state-of-the-art ap- proaches with a better comprehension of the words mean- ings. While most of the …

Content truncated to 500 words