The AFRL IWSLT 2020 Systems: Work-From-Home Edition

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 103–108 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 103 The AFRL IWSLT 2020 Systems: Work-From-Home Edition Brian Ore, Eric Hansen, Timothy Anderson, Jeremy Gwinnup Air Force Research Laboratory {brian.ore.1,eric.hansen.5, timothy.anderson.20, jeremy.gwinnup.1}@us.af.mil Abstract This report summarizes the Air Force Re- search Laboratory (AFRL) submission to the offline spoken language translation (SLT) task as part of the IWSLT 2020 evaluation cam- paign. As in previous years, we chose to adopt the cascade approach of using separate sys- tems to perform speech activity detection, au- tomatic speech recognition, sentence segmen- tation, and machine translation. All systems were neural based, including a fully-connected neural network for speech activity detection, a Kaldi factorized time delay neural network with recurrent neural network (RNN) language model rescoring for speech recognition, a bidi- rectional RNN with attention mechanism for sentence segmentation, and transformer net- works trained with OpenNMT and Marian for machine translation. Our primary submission yielded BLEU scores of 21.28 on tst2019 and 23.33 on tst2020. 1 Introduction As part of the evaluation campaign for the 2020 In- ternational Workshop on Spoken Language Trans- lation (IWSLT) (Ansari et al., 2020), the AFRL Hu- man Language Technology team submitted an en- try to the offline spoken language translation (SLT) task. The goal of this task is to automatically gener- ate cased and punctuated German translations from English audio TED Talks using either a cascade of systems or the end-to-end approach. We chose to build upon our previous work (Ore et al., 2018; Kazi et al., 2016) and adopt the cascade approach of using separate systems to perform speech activ- ity detection, automatic speech recognition (ASR), sentence segmentation, and machine translation (MT). Sections 2 and 3 describe our ASR and MT systems, respectively. Section 4 presents our re- sults on the development set when the data is man- ually segmented into sentences, and Section 5 de- scibes our approach to SLT on unsegmented data. Section 6 provides a post-evaluation analysis of our systems based on sentence length, and Section 7 presents our conclusions and future work. 2 Automatic Speech Recognition This section describes the English ASR system that was developed for the offline speech trans- lation task. First, we sequestered all talks from TEDLIUM-v3 (Hernandez et al., 2018) that were present in tst2014, tst2015, and tst2018. Next, language models (LMs) were estimated on TEDLIUM-v3 and the same subsets of News Crawl and News Discussions described in Ore et al. (2018). The text was formatted as follows: • Numbers and special symbols were converted to words (e.g., “%” converted to “percent”, “&” converted to “and”, “=” converted to “equals”). • Punctuation marks and any remaining sym- bols were removed. • All text was converted to lowercase. We used the SRILM select-vocab tool1 to choose a 100,000 word vocabulary. An interpo- lated bigram language model (LM) was estimated using the SRILM toolkit, and a recurrent neu- ral network (RNN) LM was trained using Kaldi (Povey et al., …

Content truncated to 500 words