Exploiting semantic relationships for unsupervised expansion of sentiment lexicons. (December 2020)

• Record Type: Journal Article
• Authors: Viegas, Felipe; Alvim, Mário S.; Canuto, Sérgio; Rosa, Thierson; Gonçalves, Marcos André; Rocha, Leonardo
• Is Part Of: Information systems. Volume 94(2020)
• Publication Date: 2020-12
• Subjects: Sentiment analysis -- Lexicon dictionary -- Word embeddings -- Lexicon expansion; Database management -- Periodicals; Electronic data processing -- Periodicals; Bases de données -- Gestion -- Périodiques; Informatique -- Périodiques; Database management; Electronic data processing; Periodicals; 005.7
• DOI: 10.1016/j.is.2020.101606
• Languages: English
• ISSNs: 0306-4379

Abstract

Abstract: The literature in sentiment analysis has widely assumed that semantic relationships between words cannot be effectively exploited to produce satisfactory sentiment lexicon expansions. This assumption stems from the fact that words considered to be "close" in a semantic space (e.g., word embeddings) may present completely opposite polarities, which might suggest that sentiment information in such spaces is either too faint, or at least not readily exploitable. Our main contribution in this paper is a rigorous and robust challenge to this assumption: by proposing a set of theoretical hypotheses and corroborating them with strong experimental evidence, we demonstrate that semantic relationships can be effectively used for good lexicon expansion. Based on these results, our second contribution is a novel, simple, and yet effective lexicon-expansion strategy based on semantic relationships extracted from word embeddings. This strategy is able to substantially enhance the lexicons, whilst overcoming the major problem of lexicon coverage . We present an extensive experimental evaluation of sentence-level sentiment analysis, comparing our approach to sixteen state-of-the-art (SOTA) lexicon-based and five lexicon expansion methods, over twenty datasets. Results show that in the vast majority of cases our approach outperforms the alternatives, achieving coverage of almost 100% and gains of about 26% against the best baselines. Moreover, our unsupervised approach performed competitively against SOTA supervised sentiment analysis methods, mainly in scenarios with scarce information. Finally, in a cross-dataset comparison, our approach turned out to be as competitive as (i.e., statistically tie with) state-of-the-art supervised solutions such as pre-trained transformers (BERT), even without relying on any training (labeled) data. Indeed in small datasets or in datasets with scarce information (short messages), our solution outperformed the supervised ones by large margins. Highlights: We demonstrate that semantic relationships can be effective for lexicon expansion. We propose a novel method that explores distances between word embeddings. Our unsupervised method enhances lexicon coverage while keeping high precision. In our experiments, we beat all lexicon-expansion baselines by large margins. Our approach is competitive with pre-trained transformers (BERT) without any training.