The Volctrans Neural Speech Translation System for IWSLT 2021

This paper describes the systems submitted to IWSLT 2021 by the Volctrans team. We participate in the offline speech translation and text-to-text simultaneous translation tracks. For offline speech translation, our best end-to-end model achieves 8.1 BLEU improvements over the benchmark on the MuST-C test set and is even approaching the results of a strong cascade solution. For text-to-text simultaneous translation, we explore the best practice to optimize the wait-k model. As a result, our final submitted systems exceed the benchmark at around 7 BLEU on the same latency regime. We will publish our code and model to facilitate both future research works and industrial applications.

Learning Shared Semantic Space for Speech-to-Text Translation

Having numerous potential applications and great impact, end-to-end speech translation (ST) has long been treated as an independent task, failing to fully draw strength from the rapid advances of its sibling - text machine translation (MT). With text and audio inputs represented differently, the modality gap has rendered MT data and its end-to-end models incompatible with their ST counterparts. In observation of this obstacle, we propose to bridge this representation gap with Chimera. By projecting audio and text features to a common semantic representation, Chimera unifies MT and ST tasks and boosts the performance on ST benchmark, MuST-C, to a new state-of-the-art. Specifically, Chimera obtains 26.3 BLEU on EN-DE, improving the SOTA by a +2.7 BLEU margin. Further experimental analyses demonstrate that the shared semantic space indeed conveys common knowledge between these two tasks and thus paves a new way for augmenting training resources across modalities.

Autocorrect in the Process of Translation -- Multi-task Learning Improves Dialogue Machine Translation

Automatic translation of dialogue texts is a much needed demand in many real life scenarios. However, the currently existing neural machine translation delivers unsatisfying results. In this paper, we conduct a deep analysis of a dialogue corpus and summarize three major issues on dialogue translation, including pronoun dropping (\droppro), punctuation dropping (\droppun), and typos (\typo). In response to these challenges, we propose a joint learning method to identify omission and typo, and utilize context to translate dialogue utterances. To properly evaluate the performance, we propose a manually annotated dataset with 1,931 Chinese-English parallel utterances from 300 dialogues as a benchmark testbed for dialogue translation. Our experiments show that the proposed method improves translation quality by 3.2 BLEU over the baselines. It also elevates the recovery rate of omitted pronouns from 26.09% to 47.16%. We will publish the code and dataset publicly at https://github.com/rgwt123/DialogueMT.

End-to-end Speech Translation via Cross-modal Progressive Training

End-to-end speech translation models have become a new trend in the research due to their potential of reducing error propagation. However, these models still suffer from the challenge of data scarcity. How to effectively make use of unlabeled or other parallel corpora from machine translation is promising but still an open problem. In this paper, we propose Cross Speech-Text Network (XSTNet), an end-to-end model for speech-to-text translation. XSTNet takes both speech and text as input and outputs both transcription and translation text. The model benefits from its three key design aspects: a self supervising pre-trained sub-network as the audio encoder, a multi-task training objective to exploit additional parallel bilingual text, and a progressive training procedure. We evaluate the performance of XSTNet and baselines on the MuST-C En-De/Fr/Ru datasets. XSTNet achieves state-of-the-art results on all three language directions with an average BLEU of 27.8, outperforming the previous best method by 3.7 BLEU. The code and the models will be released to the public.

Serial or Parallel? Plug-able Adapter for multilingual machine translation

Developing a unified multilingual translation model is a key topic in machine translation research. However, existing approaches suffer from performance degradation: multilingual models yield inferior performance compared to the ones trained separately on rich bilingual data. We attribute the performance degradation to two issues: multilingual embedding conflation and multilingual fusion effects. To address the two issues, we propose PAM, a Transformer model augmented with defusion adaptation for multilingual machine translation. Specifically, PAM consists of embedding and layer adapters to shift the word and intermediate representations towards language-specific ones. Extensive experiment results on IWSLT, OPUS-100, and WMT benchmarks show that \method outperforms several strong competitors, including series adapter and multilingual knowledge distillation.

Auto Correcting in the Process of Translation -- Multi-task Learning Improves Dialogue Machine Translation

Automatic translation of dialogue texts is a much needed demand in many real life scenarios. However, the currently existing neural machine translation delivers unsatisfying results. In this paper, we conduct a deep analysis of a dialogue corpus and summarize three major issues on dialogue translation, including pronoun dropping (\droppro), punctuation dropping (\droppun), and typos (\typo). In response to these challenges, we propose a joint learning method to identify omission and typo, and utilize context to translate dialogue utterances. To properly evaluate the performance, we propose a manually annotated dataset with 1,931 Chinese-English parallel utterances from 300 dialogues as a benchmark testbed for dialogue translation. Our experiments show that the proposed method improves translation quality by 3.2 BLEU over the baselines. It also elevates the recovery rate of omitted pronouns from 26.09% to 47.16%. We will publish the code and dataset publicly at https://github.com/rgwt123/DialogueMT.

Finding Sparse Structure for Domain Specific Neural Machine Translation

Fine-tuning is a major approach for domain adaptation in Neural Machine Translation (NMT). However, unconstrained fine-tuning requires very careful hyper-parameter tuning otherwise it is easy to fall into over-fitting on the target domain and degradation on the general domain. To mitigate it, we propose PRUNE-TUNE, a novel domain adaptation method via gradual pruning. It learns tiny domain-specific subnetworks for tuning. During adaptation to a new domain, we only tune its corresponding subnetwork. PRUNE-TUNE alleviates the over-fitting and the degradation problem without model modification. Additionally, with no overlapping between domain-specific subnetworks, PRUNE-TUNE is also capable of sequential multi-domain learning. Empirical experiment results show that PRUNE-TUNE outperforms several strong competitors in the target domain test set without the quality degradation of the general domain in both single and multiple domain settings.

NeurST: Neural Speech Translation Toolkit

NeurST is an open-source toolkit for neural speech translation developed by ByteDance AI Lab. The toolkit mainly focuses on end-to-end speech translation, which is easy to use, modify, and extend to advanced speech translation research and products. NeurST aims at facilitating the speech translation research for NLP researchers and provides a complete setup for speech translation benchmarks, including feature extraction, data preprocessing, distributed training, and evaluation. Moreover, The toolkit implements several major architectures for end-to-end speech translation. It shows experimental results for different benchmark datasets, which can be regarded as reliable baselines for future research. The toolkit is publicly available at https://github.com/bytedance/neurst.

Reciprocal Supervised Learning Improves Neural Machine Translation

Despite the recent success on image classification, self-training has only achieved limited gains on structured prediction tasks such as neural machine translation (NMT). This is mainly due to the compositionality of the target space, where the far-away prediction hypotheses lead to the notorious reinforced mistake problem. In this paper, we revisit the utilization of multiple diverse models and present a simple yet effective approach named Reciprocal-Supervised Learning (RSL). RSL first exploits individual models to generate pseudo parallel data, and then cooperatively trains each model on the combined synthetic corpus. RSL leverages the fact that different parameterized models have different inductive biases, and better predictions can be made by jointly exploiting the agreement among each other. Unlike the previous knowledge distillation methods built upon a much stronger teacher, RSL is capable of boosting the accuracy of one model by introducing other comparable or even weaker models. RSL can also be viewed as a more efficient alternative to ensemble. Extensive experiments demonstrate the superior performance of RSL on several benchmarks with significant margins.

On the Sentence Embeddings from Pre-trained Language Models

Pre-trained contextual representations like BERT have achieved great success in natural language processing. However, the sentence embeddings from the pre-trained language models without fine-tuning have been found to poorly capture semantic meaning of sentences. In this paper, we argue that the semantic information in the BERT embeddings is not fully exploited. We first reveal the theoretical connection between the masked language model pre-training objective and the semantic similarity task theoretically, and then analyze the BERT sentence embeddings empirically. We find that BERT always induces a non-smooth anisotropic semantic space of sentences, which harms its performance of semantic similarity. To address this issue, we propose to transform the anisotropic sentence embedding distribution to a smooth and isotropic Gaussian distribution through normalizing flows that are learned with an unsupervised objective. Experimental results show that our proposed BERT-flow method obtains significant performance gains over the state-of-the-art sentence embeddings on a variety of semantic textual similarity tasks. The code is available at https://github.com/bohanli/BERT-flow.