Evaluating Discourse Phenomena in Neural Machine Translation

Alexandra Birch
University of Edinburgh, UK

18 April 2018
2:15pm - 3:30pm
Robotarium Seminar Room
Earl Mountbatten Building


For machine translation to tackle discourse phenomena, models must have access to extra-sentential linguistic context. There has been recent interest in modelling context in neural machine translation (NMT), but models have been principally evaluated with standard automatic metrics, poorly adapted to evaluating discourse phenomena. In this talk, I present our work on creating hand-crafted, discourse test sets, designed to test the models' ability to exploit previous source and target sentences. We investigate the performance of recently proposed multi-encoder NMT models trained on subtitles for English to French. We also explore a novel way of exploiting context from the previous sentence. A simple strategy of decoding the concatenation of the previous and current sentence leads to good performance, and our novel strategy of multi-encoding and decoding of two sentences leads to the best performance (22.5% better accuracy for coreference and 3.5% better for coherence/cohesion), highlighting the importance of target-side context.


Alexandra Birch is currently a senior research associate in Informatics at the University of Edinburgh. Alexandra has been working in the field of machine translation on many different sub-fields including reordering, evaluation, semantics, and spoken language translation. Her recent interests have focussed on neural machine translation where advances using sub-word units and monolingual data have beaten state-of-the-art baselines. She is the scientific project manager for the new EU big data project, Scalable Understanding of Multilingual Media.

Host: Verena Rieser