Context-Aware Sentence Classification in Evidence-Based Medicine

Using public domain resources and off-the-shelf tools to produce high-quality multimedia texts Manny Rayner FTI/TIM University of Geneva, Switzerland Emmanuel.Rayner@unige.ch Belinda Chiera The University of South Australia Adelaide, Australia Belinda.Chiera@unisa.edu.au Cathy Chua Independent scholar Adelaide, Australia cathyc@pioneerbooks.com.au Abstract In the turbulent world of 2022, where mass population movements due to war and disas- ter are becoming increasingly common, lan- guage skills are more relevant than ever. Peo- ple who wish to achieve a high level of pro- ficiency when learning a new language bene- fit from reading literary texts, but many learn- ers find this a challenging hurdle. Annotating texts with integrated audio and translations is a popular way to try and make them easier to approach. However, doing this automatically with TTS and machine translation engines pro- duces unengaging results, while human anno- tation is slow and expensive. Here, we present a method that uses simple scripts and readily available computational resources for speech recognition and sentence alignment to combine public-domain resources from sites like Guten- berg and LibriVox into high-quality annotated multimedia versions of literary texts. Initial results with French texts of up to 80K words in length are promising, with audio/text word error rates under 0.25% and audio/translation word error rates around 1%, producing results that are usable after only minimal postediting. 1 Introduction and motivation In Anthony Powell’s semi-autobiographical WW II novel The Soldier’s Art, the narrator mentions to his division commander that he can read Balzac in the original French, and is surprised by the re- sponse: General Liddament immediately tells him to apply for a job in Military Intelligence. Since 1943, there have of course been some important changes. English has firmly established itself as the world language, and language technology has made enormous progress, but the fundamentals are the same. People with strong language skills are still prized by the security services, who see lit- tle prospect of replacing them with Google Trans- late and related AI/ML-based technologies. Large- scale movements of linguistic communities, driven by war, climate change, and economic disaster, are making these skills increasingly relevant, not just to Intelligence but to many related sectors in- cluding immigration, law enforcement and social services. Learning to read complex texts is an es- sential component in acquiring high level language skills. Duolingo and similar gamified platforms are a popular way to get started with a new language and reach low intermediate level, but they will not give the large vocabulary and grasp of idiom that comes from extensive reading. Benchmarks for language skills are competency in reading, writing, listening and speaking. A sim- ple but effective technology for supporting the de- velopment of reading skills, widely used at least since the days of the Roman Empire (Dickey, 2016), is the bilingual text: the text is divided into segments, each one paired with a gloss/translation in the annotation language. More recently, Reading While Listening (RWL; Woodall, 2010; Isozaki, 2014; Chang and Millett, 2014; Friedland et al., 2017; Pellicer-Sánchez et al., 2018; Schwieter and Benati, …

Content truncated to 500 words