Automatic Text Simplification Tools for Deaf and Hard of Hearing Adults: Benefits of Lexical Simplification and Providing Users with Autonomy
Oliver Alonzo, Matthew Seita, Abraham Glasser, Matt Huenerfauth · 2020 · Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI '20) · doi:10.1145/3313831.3376563
Summary
This CHI 2020 paper is the first empirical study of lexical Automatic Text Simplification (ATS) with Deaf and Hard-of-Hearing (DHH) adult readers, and introduces user autonomy as a central design variable in accessibility ATS tools. Prior ATS research with DHH users had looked at syntactic simplification (rewriting sentence structure) but not lexical simplification (replacing individual difficult words with simpler synonyms), even though vocabulary knowledge is a known source of reading difficulty for DHH adults with lower literacy. The authors explore two user-interface design dimensions: user initiative (how the user triggers a simplification, ranging from fully automatic to click-activated on-demand) and change visibility (how the system indicates that a replacement has happened — no trace, a background-colour highlight, a hover pop-up showing the simpler word alongside the original, or a sidebar). A preliminary formative study (n=12 DHH participants, interviews with video prototypes) narrowed the design space, and a main experimental study (n=25 DHH participants, mean age 23.5, 10 Deaf / 3 HoH / 6 deaf / 1 non-binary, 70-minute sessions in English or ASL, $40 compensation, WRAT-4 and Index of Autonomous Functioning administered) compared four conditions on four ScienceDaily articles: original (no simplification), automatic (all complex words replaced before the user sees the text), decoration (click to reveal simpler synonym, highlighted in yellow), and pop-up (hover to preview simpler word without replacing). Simplifications were Wizard-of-Oz, authored by a team of two Deaf and one non-native-English hearing researcher using Par4Sim and SimplePPDB++ as tools.
Key findings
Objective comprehension scores did not differ significantly across the four conditions (consistent with prior lexical-simplification studies that often show the subjective benefit without matching objective gains). However, the two on-demand conditions (pop-up and decoration) received significantly higher subjective 'I was able to understand this text well' judgements than the original, and significantly higher 'I would be likely to use this' ratings than the fully automatic condition (pop-up vs automatic p = 0.0046, decoration vs automatic p = 0.00049, Wilcoxon signed-rank with Bonferroni corrections). Open-ended comments clustered around three reasons DHH participants preferred on-demand: (1) a desire to choose whether to invoke assistance, (2) a desire to see the original text — important for contexts like classwork where the original terminology matters, and (3) a desire to learn new vocabulary ('I have no way of knowing what words were used when it was automatically simplified, therefore no way to learn'). Participants disliked squiggly underlines (read as error markers) and traditional hyperlink styling (confusable with links), leading the authors to recommend coloured background highlighting.
Relevance
For accessibility practitioners, this paper is foundational in three ways. First, it establishes that 'more AI assistance' is not automatically better — fully automatic simplification received the lowest acceptance scores even though it required the least user effort, because it removed the user's sense of control and their ability to learn from the original text. Second, it frames autonomy as a first-class design dimension for reading-assistance tools, extending the Deci & Ryan self-determination tradition into the accessibility literature alongside Rello et al.'s on-demand dyslexia work. Third, it gives concrete interface-level guidance: prefer coloured background highlighting over underlines, expose originals on demand, and avoid pre-modifying text. Limitations include the sample size (25), lexical-only simplification (no syntactic or hybrid comparison), Wizard-of-Oz simplifications that ducked the question of how real ATS errors would change user trust, a single article genre (science news), and the lack of reading-speed or eye-tracking measures. The methodological-research questions this paper surfaced drove the team's 2021 complexity-evaluation and 2022 fluency-evaluation papers.
Tags: automatic text simplification · deaf and hard of hearing · lexical simplification · user autonomy · reading accessibility · readability · user experience · qualitative research