From Simultaneous to Streaming Machine Translation by Leveraging Streaming History

The Streaming ST scenario presents many challenges, but there are also opportunities that can be used to improve translation quality. This work introduces the concept of Streaming history, which holds the information of the previously translated segments. The proposed MT system is able to leverage this contextual information in order to improve translation quality.

Access paper here

This is how I described this publication in my thesis:

This paper presents a general methodology for building context-aware state-of-the-art streaming MT systems, by incorporating the previously developed DS streaming segmenter and using our proposed streaming metrics for evaluation. This publication takes advantage of the insights developed in the previous publications in order to build a strong streaming baseline MT system, and improves it with a novel context-aware training methodology which obtains significant improvements. Further improvements are also obtained with a proposed Partial Bidirectional Encoder (PBE) that has access to a larger portion of the input prefix.

Our approach is similar to the concatenative approach used in context-aware MT, and uses a sliding window which contains the previous streaming history that has been produced during the translation process. History-augmented training samples are constructed from document-level corpora, and at inference time, the real streaming history is used. Extensive experiments are carried on IWSLT English-German data in order to study the behaviour of the model and optimize the latency-quality trade-off. Using our proposed streaming latency metrics, our system is compared with the ACT streaming approach (Schneider and Alexander Waibel 2020) and the submissions to the IWSLT 2020 simultaneous track (Ansari et al. 2020), achieving a similar level of quality for a fraction of the latency.