Creating an online translation platform to build target

Creating an online translation platform to build target language resources for a medical phraselator Johanna Gerlach, Hervé Spechbach, Pierrette Bouillon

Context Translation platform Evaluation Conclusion

Context 52% foreign patients Patients at Geneva University Hospitals (HUG) 12% do not speak French at all no common language with medical staff

Interpreter or phone interpreting service availability cost MT / Google Translate unreliable (Patil et al. 2014, Bouillon et al. 2017) language availability confidentiality issues Medical phraselators (Medi. Babble, Universaldoctor) usability (Boujon et al. 2017)

Babel. Dr Collaboration between the Faculty of translation and interpreting and the HUG Speech enabled phraselator for anamnesis dialogues

Babel. Dr Based on linguistic rules (synchonized context-free grammars (Rayner et al, 2016)) Billions of spoken variations mapped to thousands of canonical sentences Canonicals pre-translated by qualified translators French languages needed at HUG: Spanish, Arabic, Tigrinya, Persian, Albanian, Romanian

Il y a combien de jours que vous avez remarqué de la fièvre ? 1. Spoken question 3. 2. Validation of Speech canonical recognition + mapping to the closest canonical sentence 4. Translation shown and spoken for the patient 5. Patient responds non-verbally

Handling repetitive content: variables Compositional sentences with variables translated separately and replaced during compilation Did you take $$medication $$med_duration ? • aspirin • cortisone-based drugs • a treatment to block immune reactions • … • • for several days for two days for one week … Did you take aspirin for several weeks? Did you take cortisone-based drugs for several months ? etc.

Building target language resources

Context Translation platform Evaluation Conclusion

Online translation platform upload & create task translation revision correction download & deploy

translation Tabular layout usual for translation memories Integrated translation memory revision correction

translation revision Separate translation of sentences and variable values Translator can see sentences with variables replaced & variables in context Possibility to add new noncompositional items correction

translation revision correction Variables replaced Validation of the sentences exactly as the patient will see them Minimal expansion (each value used at least once) Online (in the platform) or offline (in a Word document) Revisor adds comments to sentences

translation revision correction Same interface as for translation: sentences and variables separate Revisor comments shown as annotations to corresponding sentences

Context Translation platform Evaluation Conclusion

Evaluation 5 translators Albanian, Arabic, Persian, Spanish & Tigrinya Volume translated ~ 2'000 sentences & 800 variable values each Anonymous questionnaire 18 questions (Likert scales) 2 Aspects: Technical Linguistic

Technical aspects – platform features All translators who had previously translated for Babel. Dr without the platform (3) found it was an improvement All translators found the variable replacement functionalities helpful overall technical difficulty (very high / average / low) average (2) low (3)

Linguistic aspects – difficulties with compositional sentences always often rarely never 5 4 3 2 1 0 segmentation lack of generic expression

Context Translation platform Evaluation Conclusion

Conclusion Translation platform facilitates creation of target language resources More linguistic than technical difficulties