AppTek’s Submission to the IWSLT 2022 Isometric Spoken Language Translation Task

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 369 - 378 May 26-27, 2022 c⃝2022 Association for Computational Linguistics AppTek’s Submission to the IWSLT 2022 Isometric Spoken Language Translation Task Patrick Wilken AppTek Aachen, Germany pwilken@apptek.com Evgeny Matusov AppTek Aachen, Germany ematusov@apptek.com Abstract To participate in the Isometric Spoken Lan- guage Translation Task of the IWSLT 2022 evaluation, constrained condition, AppTek de- veloped neural Transformer-based systems for English-to-German with various mechanisms of length control, ranging from source-side and target-side pseudo-tokens to encoding of re- maining length in characters that replaces po- sitional encoding. We further increased trans- lation length compliance by sentence-level se- lection of length-compliant hypotheses from different system variants, as well as rescor- ing of N-best candidates from a single system. Length-compliant back-translated and forward- translated synthetic data, as well as other par- allel data variants derived from the original MuST-C training corpus were important for a good quality/desired length trade-off. Our experimental results show that length compli- ance levels above 90% can be reached while minimizing losses in MT quality as measured in BERT and BLEU scores. 1 Introduction In this paper, we describe AppTek’s submission to the IWSLT 2022 Isometric Spoken Language Translation evaluation (Anastasopoulos et al., 2022). Our goal was to create a system that pro- duces translations which are within 10% of the source sentence length, but have similar levels of quality as a baseline system translations without length control. AppTek participated in the con- strained condition with an English-to-German neu- ral machine translation (NMT) system that we de- scribe in Section 2. The system was extended with 5 different length control methods, which we explain in detail in Section 3. We also cre- ated synthetic data with back-translation, forward- translation, as well as a novel data augmentation method of synonym replacement. All three meth- ods are described in Section 4. Our experimental results on the MuST-C tst-COMMON test set and the official evaluation test set are presented in Sec- tion 5, including ablation studies that prove the effectiveness of synthetic data and noisy length en- coding for a better trade-off between length compli- ance and MT quality. We summarize our findings in Section 6. 2 Baseline system 2.1 Data We follow the constrained condition of the IWSLT Isometric SLT task and use only English-to- German TED-talk data from the MuST-C corpus (Di Gangi et al., 2019). The corpus contains 251K sentence pairs with 4.7M and 4.3M English and German words, respectively. We apply minimal text pre-processing, mainly consisting of normalization of quotes and dashes. 2K sentences that have mismatching digits or paren- theses in source and target were filtered out. We use a joint English and German Sentence- Piece model (Kudo and Richardson, 2018), trained on the whole corpus using a vocabulary size of 20K, to split the data into subwords. 2.2 Neural NMT model In preliminary experiments we tried several Trans- former model configurations, including base and big from the original paper (Vaswani et al., 2017), a …

Content truncated to 500 words