Towards Readability-Controlled Machine Translation of COVID-19 Texts

Towards Readability-Controlled Machine Translation of COVID-19 Texts Fernando Alva-Manchego Cardiff University alvamanchegof@cardiff.ac.uk Matthew Shardlow Manchester Metropolitan University m.shardlow@mmu.ac.uk Abstract This project investigates the capabilities of machine translation (MT) models for generating translations at varying levels of readability, focusing on texts about COVID-19. Funded by the European As- sociation for Machine Translation and by the Centre for Advanced Computational Sciences at Manchester Metropolitan Uni- versity, we collected manual simplifica- tions for English and Spanish texts in the TICO-19 dataset, and assessed the perfor- mance of neural MT models in this new benchmark. Future work will implement models that jointly translate and simplify, and develop suitable evaluation metrics. 1 Introduction “Multilingual Translation with Readability- Controlled Output Generation” is a project that received funding from the European Association for Machine Translation (under its programme “2021 Sponsorship of Activities”) and from the Centre for Advanced Computational Sciences at Manchester Metropolitan University. We aim to develop machine translation (MT) models that generate translations that can be understood by non-expert readers, focusing on texts with medical information. This is pertinent in the context of the COVID-19 pandemic, where there is a disparity in the availability of health-related content produced in English, compared to other languages. The project has the following objectives: (1) to collect a dataset with simplified versions of paral- lel texts in English and Spanish about COVID-19; (2) to assess how well existing state-of-the-art MT models perform on our new benchmark; and (3) to © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. Lang. Complexity W/S Sy/W FRE↑ S-P↑ English Original 23.015 6.444 45.69 – Simplified 21.838 6.308 52.70 – Spanish Original 27.623 6.287 – 75.17 Simplified 24.749 6.271 – 79.49 Table 1: Statistics of Simple TICO-19: average number of words per sentence (W/S), average number of syllables per word (Sy/W), and estimated readability with Flesch Reading Ease (FRE) for English and Szigriszt-Pazos (S-P) for Spanish. investigate additional model architectures and/or resources that are needed to generate and evaluate simplified in-domain translations. The first two goals of the project were carried out from January 2021 to December 2021, and re- sulted in the release of the Simple TICO-19 dataset (Shardlow and Alva-Manchego, 2022).1 We con- tinue to work with the new dataset to further in- vestigate the nature of readability-controlled out- put generation in the MT context. We hope to ap- ply for further funding at a national and European level as a result of this work. 2 The Simple TICO-19 Dataset We leveraged the TICO-19 benchmark (Anasta- sopoulos et al., 2020), which contains 3,000 sen- tences related to the COVID-19 pandemic, trans- lated from English into 36 languages and from sev- eral sources (e.g. academic publications, speech corpora, news articles, etc.). For our project, we collected manual simplifications for the English and Spanish subsets, resulting in the Simple TICO- 19 dataset, where each sentence has either a sim- plified version of itself, or a decision has been taken that the sentence is …

Content truncated to 500 words