Johanna Gerlach, Hervé Spechbach, Pierrette Bouillon 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
Patients at Geneva University Hospitals (HUG) 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 phraselator : based on pretranslated sentences, navigation in menu Medical phraselators (MediBabble, Universaldoctor) usability (Boujon et al. 2017)
BabelDr Collaboration between the Faculty of translation and interpreting and the HUG Speech enabled phraselator for anamnesis dialogues Phraselator -> pre-translated set of sentences Improvement : speech instead of menus
BabelDr 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
+ mapping to the closest canonical sentence 3. Validation of canonical Il y a combien de jours que vous avez remarqué de la fièvre ? 1. Spoken question 2. Speech recognition + mapping to the closest canonical sentence 3. Validation of canonical 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 aspirin for several weeks? 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 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 revision correction Tabular layout usual for translation memories Integrated translation memory
translation revision correction Separate translation of sentences and variable values Translator can see sentences with variables replaced & variables in context Possibility to add new non- compositional items
Online (in the platform) or offline (in a Word document) 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 Volume translated Anonymous questionnaire 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 BabelDr without the platform (3) found it was an improvement All translators found the variable replacement functionalities helpful average (2) low (3)
Linguistic aspects – difficulties with compositional sentences We asked the translators how often they encountered difficulties with the translation of compositional sentences : segmentation : can it easily be transposed to target language ? generic expression : is corresponding generic expression (e.g. verb) available in target language ?
Context Translation platform Evaluation Conclusion
Conclusion Translation platform facilitates creation of target language resources More linguistic than technical difficulties