“Was it “stated” or was it “claimed”?: How linguistic bias affects generative language models

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10080–10095 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 10080 Was it “said” or was it “claimed”? How linguistic bias affects generative language models Roma Patel Brown University romapatel@brown.edu Ellie Pavlick Brown University ellie pavlick@brown.edu Abstract People use language in subtle and nuanced ways to convey their beliefs. For instance, saying claimed instead of said casts doubt on the truthfulness of the underlying propo- sition, thus representing the author’s opinion on the matter. Several works have identified classes of words that induce such framing ef- fects. In this paper, we test whether gener- ative language models are sensitive to these linguistic cues. In particular, we test whether prompts that contain linguistic markers of au- thor bias (e.g., hedges, implicatives, subjec- tive intensifiers, assertives) influence the dis- tribution of the generated text. Although these framing effects are subtle and stylistic, we find qualitative and quantitative evidence that they lead to measurable style and topic differences in the generated text, leading to language that is more polarised (both positively and nega- tively) and, anecdotally, appears more skewed towards controversial entities and events. 1 Introduction With subtle changes in word choice, a writer can in- fluence a reader’s perspective on a matter in many ways (Thomas et al., 2006; Recasens et al., 2013). For example, Table 1 shows how the verbs claimed and said, although reasonable paraphrases for one another in the given sentence, have very different implications. Saying claimed casts doubt on the certainty of the underlying proposition and might implicitly bias a reader’s interpretation of the sen- tence. That is, such linguistic cues (e.g., hedges, implicatives, intensifiers) can induce subtle biases through implied sentiment and presupposed facts about the entities and events with which they in- teract (Rashkin et al., 2015). When models of lan- guage are trained on large web corpora that consist of text written by many people, distributional pat- terns might lead the lexical representations of these Bias Prompt (Assertive) Neutral Prompt In a speech on June 9, 2005, Bush claimed that the “Pa- triot Act” had been used to bring charges against more than 400 suspects, more than half of whom had been convicted. William Graff, a former Texas primary voter who was also shot on his go- go days, was shot and killed at one point in the fight be- tween Bush and the two ter- rorists, which Bush called executive order had taken “adrenaline.” In a speech on June 9, 2005, Bush said that the “Patriot Act” had been used to bring charges against more than 400 suspects, more than half of whom had been convicted. “This agreement done are out of a domestic legal order,” Bush said in referring to the presidential Domestic Vio- lence policy and the pres- ident’s new domestic vio- lence policy; Roe v. Wade. “The president is calling on everyone.. Table 1: Table shows generations from a language model (GPT-2); when prompted with a linguistically biased sentence …

Content truncated to 500 words