Unsupervised Vocabulary Selection for Domain-Independent Simultaneous Lecture Translation

In this work, we investigate methods to automatically adapt our simultaneous lecture translation systems to the diverse topics that occur in educational lectures. Utilizing materials that are available before the lecture begins, such as lecture slides, our proposed framework iteratively searches for related documents on the World Wide Web and generates lecture-specific models and vocabularies based on the resulting documents. In this paper, we propose a novel method for vocabulary selection, a critical aspect of simultaneous translation systems where the occurrence of out-of-vocabulary words significantly degrades intelligibility. We propose a novel approach based on feature-based ranking and evaluate the effectiveness of 21 different features and their combinations for this task. On the interACT German-English simultaneous lecture translation system our proposed approach significantly improved vocabulary coverage, reducing out-of-vocabulary rate, on average by 60% and up to 84%, compared to a lecture-independent baseline. Furthermore, a 40k vocabulary selected using our method obtained better coverage than a lecture-independent 300k vocabulary, improving intelligibility and reducing the latency of the end-to-end system.