This paper describes the submission of LMU Munich to the WMT 2021 multilingual machine translation task for small track #1, which studies translation between 6 languages (Croatian, Hungarian, Estonian, Serbian, Macedonian, English) in 30 directions. We investigate the extent to which bilingual translation systems can influence multilingual translation systems. More specifically, we trained 30 bilingual translation systems, covering all language pairs, and used data augmentation technologies such as back-translation and knowledge distillation to improve the multilingual translation systems. Our best translation system scores 5 to 6 BLEU higher than a strong baseline system provided by the organizers. As seen in the dynalab leaderboard, our submission is the only fully constrained submission that uses only the corpus provided by the organizers and does not use any pre-trained models.

We present the findings of the WMT2021 Shared Tasks in Unsupervised MT and Very Low Resource Supervised MT. Within the task, the community studied very low resource translation between German and Upper Sorbian, unsupervised translation between German and Lower Sorbian and low resource translation between Russian and Chuvash, all minority languages with active language communities working on preserving the languages, who are partners in the evaluation. Thanks to this, we were able to obtain most digital data available for these languages and offer them to the task participants. In total, six teams participated in the shared task. The paper discusses the background, presents the tasks and results, and discusses best practices for the future.

We present our submissions to the WMT21 shared task in Unsupervised and Very Low Resource machine translation between German and Upper Sorbian, German and Lower Sorbian, and Russian and Chuvash. Our low-resource systems (German↔Upper Sorbian, Russian↔Chuvash) are pre-trained on high-resource pairs of related languages. We fine-tune those systems using the available authentic parallel data and improve by iterated back-translation. The unsupervised German↔Lower Sorbian system is initialized by the best Upper Sorbian system and improved by iterated back-translation using monolingual data only.

Applying the Transformer architecture on the character level usually requires very deep architectures that are difficult and slow to train. These problems can be partially overcome by incorporating a segmentation into tokens in the model. We show that by initially training a subword model and then finetuning it on characters, we can obtain a neural machine translation model that works at the character level without requiring token segmentation. We use only the vanilla 6-layer Transformer Base architecture. Our character-level models better capture morphological phenomena and show more robustness to noise at the expense of somewhat worse overall translation quality. Our study is a significant step towards high-performance and easy to train character-based models that are not extremely large.

We present our systems for the WMT20 Very Low Resource MT Task for translation between German and Upper Sorbian. For training our systems, we generate synthetic data by both back- and forward-translation. Additionally, we enrich the training data with German-Czech translated from Czech to Upper Sorbian by an unsupervised statistical MT system incorporating orthographically similar word pairs and transliterations of OOV words. Our best translation system between German and Sorbian is based on transfer learning from a Czech-German system and scores 12 to 13 BLEU higher than a baseline system built using the available parallel data only.

Multilingual contextual embeddings, such as multilingual BERT and XLM-RoBERTa, have proved useful for many multi-lingual tasks. Previous work probed the cross-linguality of the representations indirectly using zero-shot transfer learning on morphological and syntactic tasks. We instead investigate the language-neutrality of multilingual contextual embeddings directly and with respect to lexical semantics. Our results show that contextual embeddings are more language-neutral and, in general, more informative than aligned static word-type embeddings, which are explicitly trained for language neutrality. Contextual embeddings are still only moderately language-neutral by default, so we propose two simple methods for achieving stronger language neutrality: first, by unsupervised centering of the representation for each language and second, by fitting an explicit projection on small parallel data. Besides, we show how to reach state-of-the-art accuracy on language identification and match the performance of statistical methods for word alignment of parallel sentences without using parallel data.

We present our submission to the Simultaneous Translation And Paraphrase for Language Education (STAPLE) challenge. We used a standard Transformer model for translation, with a crosslingual classifier predicting correct translations on the output n-best list. To increase the diversity of the outputs, we used additional data to train the translation model, and we trained a paraphrasing model based on the Levenshtein Transformer architecture to generate further synonymous translations. The paraphrasing results were again filtered using our classifier. While the use of additional data and our classifier filter were able to improve results, the paraphrasing model produced too many invalid outputs to further improve the output quality. Our model without the paraphrasing component finished in the middle of the field for the shared task, improving over the best baseline by a margin of 10-22 % weighted F1 absolute.

In this paper, we study abstractive summarization for open-domain videos. Unlike the traditional text news summarization, the goal is less to “compress” text information but rather to provide a fluent textual summary of information that has been collected and fused from different source modalities, in our case video and audio transcripts (or text). We show how a multi-source sequence-to-sequence model with hierarchical attention can integrate information from different modalities into a coherent output, compare various models trained with different modalities and present pilot experiments on the How2 corpus of instructional videos. We also propose a new evaluation metric (Content F1) for abstractive summarization task that measures semantic adequacy rather than fluency of the summaries, which is covered by metrics like ROUGE and BLEU.

We present our submission to the WMT19 Robustness Task. Our baseline system is the Charles University (CUNI) Transformer system trained for the WMT18 shared task on News Translation. Quantitative results show that the CUNI Transformer system is already far more robust to noisy input than the LSTM-based baseline provided by the task organizers. We further improved the performance of our model by fine-tuning on the in-domain noisy data without influencing the translation quality on the news domain.

Autoregressive decoding is the only part of sequence-to-sequence models that prevents them from massive parallelization at inference time. Non-autoregressive models enable the decoder to generate all output symbols independently in parallel. We present a novel non-autoregressive architecture based on connectionist temporal classification and evaluate it on the task of neural machine translation. Unlike other non-autoregressive methods which operate in several steps, our model can be trained end-to-end. We conduct experiments on the WMT English-Romanian and English-German datasets. Our models achieve a significant speedup over the autoregressive models, keeping the translation quality comparable to other non-autoregressive models.

In multi-source sequence-to-sequence tasks, the attention mechanism can be modeled in several ways. This topic has been thoroughly studied on recurrent architectures. In this paper, we extend the previous work to the encoder-decoder attention in the Transformer architecture. We propose four different input combination strategies for the encoder-decoder attention: serial, parallel, flat, and hierarchical. We evaluate our methods on tasks of multimodal translation and translation with multiple source languages. The experiments show that the models are able to use multiple sources and improve over single source baselines.

We present our submission to the WMT18 Multimodal Translation Task. The main feature of our submission is applying a self-attentive network instead of a recurrent neural network. We evaluate two methods of incorporating the visual features in the model: first, we include the image representation as another input to the network; second, we train the model to predict the visual features and use it as an auxiliary objective. For our submission, we acquired both textual and multimodal additional data. Both of the proposed methods yield significant improvements over recurrent networks and self-attentive textual baselines.

Modeling attention in neural multi-source sequence-to-sequence learning remains a relatively unexplored area, despite its usefulness in tasks that incorporate multiple source languages or modalities. We propose two novel approaches to combine the outputs of attention mechanisms over each source sequence, flat and hierarchical. We compare the proposed methods with existing techniques and present results of systematic evaluation of those methods on the WMT16 Multimodal Translation and Automatic Post-editing tasks. We show that the proposed methods achieve competitive results on both tasks.

Continuous word representations appeared to be a useful feature in many natural language processing tasks. Using fixed-dimension pre-trained word embeddings allows avoiding sparse bag-of-words representation and to train models with fewer parameters. In this paper, we use fixed pre-trained word embeddings as additional features for a neural scoring function in the MST parser. With the multi-layer architecture of the scoring function we can avoid handcrafting feature conjunctions. The continuous word representations on the input also allow us to reduce the number of lexical features, make the parser more robust to out-of-vocabulary words, and reduce the total number of parameters of the model. Although its accuracy stays below the state of the art, the model size is substantially smaller than with the standard features set. Moreover, it performs well for languages where only a smaller treebank is available and the results promise to be useful in cross-lingual parsing.