Does BERT Learn as Humans Perceive? Understanding Linguistic Styles through Lexica

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6323–6331 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 6323 Does BERT Learn as Humans Perceive? Understanding Linguistic Styles through Lexica Shirley Anugrah Hayati∗ Dongyeop Kang Lyle Ungar University of Pennsylvania Georgia Institute of Technology University of Minnesota shirley@gatech.edu dongyeop@umn.edu ungar@cis.upenn.edu Abstract People convey their intention and attitude through linguistic styles of the text that they write. In this study, we investigate lexicon usages across styles throughout two lenses: human perception and machine word impor- tance, since words differ in the strength of the stylistic cues that they provide. To col- lect labels of human perception, we curate a new dataset, HUMMINGBIRD, on top of bench- marking style datasets. We have crowd work- ers highlight the representative words in the text that makes them think the text has the following styles: politeness, sentiment, offen- siveness, and five emotion types. We then compare these human word labels with word importance derived from a popular fine-tuned style classifier like BERT. Our results show that the BERT often finds content words not relevant to the target style as important words used in style prediction, but humans do not perceive the same way even though for some styles (e.g., positive sentiment and joy) human- and machine-identified words share significant overlap for some styles.1 1 Introduction To express their interpersonal goal and attitude, people often use different styles in their communi- cation. The style of a text can be as important as its literal meaning for effective communication (Hovy, 1987). NLP researchers have built many models to identify different styles in text, including politeness (Danescu-Niculescu-Mizil et al., 2013), emotion (Alm et al., 2005; Mohammad et al., 2018), and sen- timent (Socher et al., 2013). Recently, transformer- based (Vaswani et al., 2017) pretrained language models, such as BERT (Devlin et al., 2019), have achieved impressive performance on many NLP tasks, including stylistic studies. However, explain- ing what these deep learning models learn remains ∗ research conducted at the University of Pennsylvania 1Our dataset and code are available at https:// github.com/sweetpeach/hummingbird I will understand if you decline, but would very much like you to accept. May I nominate you? (a) Human: Polite BERT: Polite a nightmare date with a half-formed wit done a great disservice by a lack of critical distance and a sad trust in liberal arts college bumper sticker platitudes . (b) Human: Anger BERT: Not Anger Human BERT Both Figure 1: Both humans and BERT models label the sentence (a) as “polite”, whereas in sentence (b), the humans label it as “anger” but BERT does not. Pink highlight: high human perception score. Blue: BERT’s important words. Purple: the word is seen as a strong cue by both human and BERT. The darker color means that the score for human perception or machine word importance is higher. Best seen in color. a challenge. Thus, there is a growing effort to un- derstand how these models behave (Rogers et al., 2021; Rajagopal et …

Content truncated to 500 words