The pre-training of masked language models (MLMs) consumes massive computation to achieve good results on downstream NLP tasks, resulting in a large carbon footprint. In the vanilla MLM, the virtual tokens, [MASK]s, act as placeholders and gather the contextualized information from unmasked tokens to restore the corrupted information. It raises the question of whether we can append [MASK]s at a later layer, to reduce the sequence length for earlier layers and make the pre-training more efficient. We show: (1) [MASK]s can indeed be appended at a later layer, being disentangled from the word embedding; (2) The gathering of contextualized information from unmasked tokens can be conducted with a few layers. By further increasing the masking rate from 15% to 50%, we can pre-train RoBERTa-base and RoBERTa-large from scratch with only 78% and 68% of the original computational budget without any degradation on the GLUE benchmark. When pre-training with the original budget, our method outperforms RoBERTa for 6 out of 8 GLUE tasks, on average by 0.4%.

This paper illustrates our approach to the shared task on large-scale multilingual machine translation in the sixth conference on machine translation (WMT-21). In this work, we aim to build a single multilingual translation system with a hypothesis that a universal cross-language representation leads to better multilingual translation performance. We extend the exploration of different back-translation methods from bilingual translation to multilingual translation. Better performance is obtained by the constrained sampling method, which is different from the finding of the bilingual translation. Besides, we also explore the effect of vocabularies and the amount of synthetic data. Surprisingly, the smaller size of vocabularies perform better, and the extensive monolingual English data offers a modest improvement. We submitted to both the small tasks and achieve the second place.

In this paper, we empirically investigate applying word-level weights to adapt neural machine translation to e-commerce domains, where small e-commerce datasets and large out-of-domain datasets are available. In order to mine in-domain like words in the out-of-domain datasets, we compute word weights by using a domain-specific and a non-domain-specific language model followed by smoothing and binary quantization. The baseline model is trained on mixed in-domain and out-of-domain datasets. Experimental results on En → Zh e-commerce domain translation show that compared to continuing training without word weights, it improves MT quality by up to 3.11% BLEU absolute and 1.59% TER. We have also trained models using fine-tuning on the in-domain data. Pre-training a model with word weights improves fine-tuning up to 1.24% BLEU absolute and 1.64% TER, respectively.

Conversational spoken language translation (CSLT) systems facilitate bilingual conversations in which the two participants speak different languages. Bilingual conversations provide additional contextual information that can be used to improve the underlying machine translation system. In this paper, we describe a novel translation model adaptation method that anticipates a participant’s response in the target language, based on his counterpart’s prior turn in the source language. Our proposed strategy uses the source language utterance to perform cross-language retrieval on a large corpus of bilingual conversations in order to obtain a set of potentially relevant target responses. The responses retrieved are used to bias translation choices towards anticipated responses. On an Iraqi-to-English CSLT task, our method achieves a significant improvement over the baseline system in terms of BLEU, TER and METEOR metrics.

We describe a novel two-way speech-to-speech (S2S) translation system that actively detects a wide variety of common error types and resolves them through user-friendly dialog with the user(s). We present algorithms for detecting out-of-vocabulary (OOV) named entities and terms, sense ambiguities, homophones, idioms, ill-formed input, etc. and discuss novel, interactive strategies for recovering from such errors. We also describe our approach for prioritizing different error types and an extensible architecture for implementing these decisions. We demonstrate the efficacy of our system by presenting analysis on live interactions in the English-to-Iraqi Arabic direction that are designed to invoke different error types for spoken language translation. Our analysis shows that the system can successfully resolve 47% of the errors, resulting in a dramatic improvement in the transfer of problematic concepts.

Phrase alignment is a crucial step in phrase-based statistical machine translation. We explore a way of improving phrase alignment by adding syntactic information in the form of chunks as soft constraints guided by an in-depth and detailed analysis on a hand-aligned data set. We extend a probabilistic phrase alignment model that extracts phrase pairs by optimizing phrase pair boundaries over the sentence pair [1]. The boundaries of the target phrase are chosen such that the overall sentence alignment probability is optimal. Viterbi alignment information is also added in the extended model with a view of improving phrase alignment. We extract phrase pairs using a relatively larger number of features which are discriminatively trained using a large-margin online learning algorithm, i.e., Margin Infused Relaxed Algorithm (MIRA) and integrate it in our approach. Initial experiments show improvements in both phrase alignment and translation quality for Arabic-English on a moderate-size translation task.