Unsupervised Vocabulary Selection for Domain-Independent Simultaneous Lecture Translation

In this work, we investigate methods to au-tomatically adapt our simultaneous lecture translation systems to the diverse topics that occur in educational lectures. Utilizing ma-terials that are available before the lecture begins, such as lecture slides, our proposed framework iteratively searches for related doc-uments on the World Wide Web and gen-erates lecture-specific models and vocabular-ies based on the resulting documents. In this paper, we propose a novel method for vocabulary selection, a critical aspect of si-multaneous translation systems where the oc-currence of out-of-vocabulary words signifi-cantly degrades intelligibility. We propose a novel approach based on feature-based rank-ing and evaluate the effectiveness of 21 dif-ferent features and their combinations for this task. On the interACT German-English si-multaneous lecture translation system our pro-posed approach significantly improved vocab-ulary coverage, reducing out-of-vocabulary rate, on average by 60 % and up to 84%, com-pared to a lecture-independent baseline. Fur-thermore, a 40k vocabulary selected using our method obtained better coverage than a lecture-independent 300k vocabulary, improv-ing intelligibility and reducing the latency of the end-to-end system.