AutoMeTS: The Autocomplete for Medical Text Simplification

Proceedings of the 28th International Conference on Computational Linguistics, pages 1424–1434 Barcelona, Spain (Online), December 8-13, 2020 1424 AutoMeTS: The Autocomplete for Medical Text Simplification Hoang Van, David Kauchak*, Gondy Leroy University of Arizona, *Pomona College {vnhh,gondyleroy}@email.arizona.edu, david.kauchak@pomona.edu Abstract The goal of text simplification (TS) is to transform difficult text into a version that is easier to understand and more broadly accessible to a wide variety of readers. In some domains, such as healthcare, fully automated approaches cannot be used since information must be accurately pre- served. Instead, semi-automated approaches can be used that assist a human writer in simplifying text faster and at a higher quality. In this paper, we examine the application of autocomplete to text simplification in the medical domain. We introduce a new parallel medical data set consisting of aligned English Wikipedia with Simple English Wikipedia sentences and examine the applica- tion of pretrained neural language models (PNLMs) on this dataset. We compare four PNLMs (BERT, RoBERTa, XLNet, and GPT-2), and show how the additional context of the sentence to be simplified can be incorporated to achieve better results (6.17% absolute improvement over the best individual model). We also introduce an ensemble model that combines the four PNLMs and outperforms the best individual model by 2.1%, resulting in an overall word prediction accuracy of 64.52%. 1 Introduction The goal of text simplification is to transform text into a variant that is more broadly accessible to a wide variety of readers while preserving the content. While this has been accomplished using a range of approaches (Shardlow, 2014), most text simplification research has focused on fully-automated approaches (Xu et al., 2016; Zhang and Lapata, 2017; Nishihara et al., 2019). However, in some domains, such as healthcare, using fully-automated text simplification is not appropriate since it is critical that the important information is preserved fully during the simplification process. For example, Shardlow et al. (2019) found that fully automated approaches omitted 30% of critical information when used to simplify clinical texts. For these types of domains, instead of fully-automated approaches, interactive text simplification tools are better suited to generate more efficient and higher quality simplifications (Kloehn et al., 2018). Autocomplete tools suggest one or more words during text composition that could follow what has been typed so far. They have been used in a range of applications including web queries (Cai et al., 2016), database queries (Khoussainova et al., 2010), texting (Dunlop and Crossan, 2000), e-mail composition (Chen et al., 2019), and interactive machine translation, where a user translating a foreign sentence is given guidance as they type (Green et al., 2014). Our work is most similar to interactive machine translation. Autocomplete tools can speed up the text simplification process and give full control over information preservation to users, which is required in some domains, such as health and medical. In this paper, we explore the application of pretrained neural language models (PNLMs) to the auto- complete process for sentence-level medical text simplification. Specifically, given (a) a difficult sentence …

Content truncated to 500 words