On Sparsifying Encoder Outputs in Sequence-to-Sequence Models

Sequence-to-sequence models usually transfer all encoder outputs to the decoder for generation. In this work, by contrast, we hypothesize that these encoder outputs can be compressed to shorten the sequence delivered for decoding. We take Transformer as the testbed and introduce a layer of stochastic gates in-between the encoder and the decoder. The gates are regularized using the expected value of the sparsity-inducing L0penalty, resulting in completely masking-out a subset of encoder outputs. In other words, via joint training, the L0DROP layer forces Transformer to route information through a subset of its encoder states. We investigate the effects of this sparsification on two machine translation and two summarization tasks. Experiments show that, depending on the task, around 40-70% of source encodings can be pruned without significantly compromising quality. The decrease of the output length endows L0DROP with the potential of improving decoding efficiency, where it yields a speedup of up to 1.65x on document summarization tasks against the standard Transformer. We analyze the L0DROP behaviour and observe that it exhibits systematic preferences for pruning certain word types, e.g., function words and punctuation get pruned most. Inspired by these observations, we explore the feasibility of specifying rule-based patterns that mask out encoder outputs based on information such as part-of-speech tags, word frequency and word position.

A Set of Recommendations for Assessing Human-Machine Parity in Language Translation

The quality of machine translation has increased remarkably over the past years, to the degree that it was found to be indistinguishable from professional human translation in a number of empirical investigations. We reassess Hassan et al.'s 2018 investigation into Chinese to English news translation, showing that the finding of human-machine parity was owed to weaknesses in the evaluation design - which is currently considered best practice in the field. We show that the professional human translations contained significantly fewer errors, and that perceived quality in human evaluation depends on the choice of raters, the availability of linguistic context, and the creation of reference translations. Our results call for revisiting current best practices to assess strong machine translation systems in general and human-machine parity in particular, for which we offer a set of recommendations based on our empirical findings.

X-Stance: A Multilingual Multi-Target Dataset for Stance Detection

We extract a large-scale stance detection dataset from comments written by candidates of elections in Switzerland. The dataset consists of German, French and Italian text, allowing for a cross-lingual evaluation of stance detection. It contains 67 000 comments on more than 150 political issues (targets). Unlike stance detection models that have specific target issues, we use the dataset to train a single model on all the issues. To make learning across targets possible, we prepend to each instance a natural question that represents the target (e.g. "Do you support X?"). Baseline results from multilingual BERT show that zero-shot cross-lingual and cross-target transfer of stance detection is moderately successful with this approach.

Domain Robustness in Neural Machine Translation

Translating text that diverges from the training domain is a key challenge for neural machine translation (NMT). Domain robustness - the generalization of models to unseen test domains - is low compared to statistical machine translation. In this paper, we investigate the performance of NMT on out-of-domain test sets, and ways to improve it. We observe that hallucination (translations that are fluent but unrelated to the source) is common in out-of-domain settings, and we empirically compare methods that improve adequacy (reconstruction), out-of-domain translation (subword regularization), or robustness against adversarial examples (defensive distillation), as well as noisy channel models. In experiments on German to English OPUS data, and German to Romansh, a low-resource scenario, we find that several methods improve domain robustness, reconstruction standing out as a method that not only improves automatic scores, but also shows improvements in a manual assessments of adequacy, albeit at some loss in fluency. However, out-of-domain performance is still relatively low and domain robustness remains an open problem.

Domain, Translationese and Noise in Synthetic Data for Neural Machine Translation

The quality of neural machine translation can be improved by leveraging additional monolingual resources to create synthetic training data. Source-side monolingual data can be (forward-)translated into the target language for self-training; target-side monolingual data can be back-translated. It has been widely reported that back-translation delivers superior results, but could this be due to artefacts in the test sets? We perform a case study using French-English news translation task and separate test sets based on their original languages. We show that forward translation delivers superior gains in terms of BLEU on sentences that were originally in the source language, complementing previous studies which show large improvements with back-translation on sentences that were originally in the target language. To better understand when and why forward and back-translation are effective, we study the role of domains, translationese, and noise. While translationese effects are well known to influence MT evaluation, we also find evidence that news data from different languages shows subtle domain differences, which is another explanation for varying performance on different portions of the test set. We perform additional low-resource experiments which demonstrate that forward translation is more sensitive to the quality of the initial translation system than back-translation, and tends to perform worse in low-resource settings.

Root Mean Square Layer Normalization

Layer normalization (LayerNorm) has been successfully applied to various deep neural networks to help stabilize training and boost model convergence because of its capability in handling re-centering and re-scaling of both inputs and weight matrix. However, the computational overhead introduced by LayerNorm makes these improvements expensive and significantly slows the underlying network, e.g. RNN in particular. In this paper, we hypothesize that re-centering invariance in LayerNorm is dispensable and propose root mean square layer normalization, or RMSNorm. RMSNorm regularizes the summed inputs to a neuron in one layer according to root mean square (RMS), giving the model re-scaling invariance property and implicit learning rate adaptation ability. RMSNorm is computationally simpler and thus more efficient than LayerNorm. We also present partial RMSNorm, or pRMSNorm where the RMS is estimated from p% of the summed inputs without breaking the above properties. Extensive experiments on several tasks using diverse network architectures show that RMSNorm achieves comparable performance against LayerNorm but reduces the running time by 7%~64% on different models. Source code is available at https://github.com/bzhangGo/rmsnorm.

Context-Aware Monolingual Repair for Neural Machine Translation

Modern sentence-level NMT systems often produce plausible translations of isolated sentences. However, when put in context, these translations may end up being inconsistent with each other. We propose a monolingual DocRepair model to correct inconsistencies between sentence-level translations. DocRepair performs automatic post-editing on a sequence of sentence-level translations, refining translations of sentences in context of each other. For training, the DocRepair model requires only monolingual document-level data in the target language. It is trained as a monolingual sequence-to-sequence model that maps inconsistent groups of sentences into consistent ones. The consistent groups come from the original training data; the inconsistent groups are obtained by sampling round-trip translations for each isolated sentence. We show that this approach successfully imitates inconsistencies we aim to fix: using contrastive evaluation, we show large improvements in the translation of several contextual phenomena in an English-Russian translation task, as well as improvements in the BLEU score. We also conduct a human evaluation and show a strong preference of the annotators to corrected translations over the baseline ones. Moreover, we analyze which discourse phenomena are hard to capture using monolingual data only.

The Bottom-up Evolution of Representations in the Transformer: A Study with Machine Translation and Language Modeling Objectives

We seek to understand how the representations of individual tokens and the structure of the learned feature space evolve between layers in deep neural networks under different learning objectives. We focus on the Transformers for our analysis as they have been shown effective on various tasks, including machine translation (MT), standard left-to-right language models (LM) and masked language modeling (MLM). Previous work used black-box probing tasks to show that the representations learned by the Transformer differ significantly depending on the objective. In this work, we use canonical correlation analysis and mutual information estimators to study how information flows across Transformer layers and how this process depends on the choice of learning objective. For example, as you go from bottom to top layers, information about the past in left-to-right language models gets vanished and predictions about the future get formed. In contrast, for MLM, representations initially acquire information about the context around the token, partially forgetting the token identity and producing a more generalized token representation. The token identity then gets recreated at the top MLM layers.

Encoders Help You Disambiguate Word Senses in Neural Machine Translation

Neural machine translation (NMT) has achieved new state-of-the-art performance in translating ambiguous words. However, it is still unclear which component dominates the process of disambiguation. In this paper, we explore the ability of NMT encoders and decoders to disambiguate word senses by evaluating hidden states and investigating the distributions of self-attention. We train a classifier to predict whether a translation is correct given the representation of an ambiguous noun. We find that encoder hidden states outperform word embeddings significantly which indicates that encoders adequately encode relevant information for disambiguation into hidden states. In contrast to encoders, the effect of decoder is different in models with different architectures. Moreover, the attention weights and attention entropy show that self-attention can detect ambiguous nouns and distribute more attention to the context.

Improving Deep Transformer with Depth-Scaled Initialization and Merged Attention

The general trend in NLP is towards increasing model capacity and performance via deeper neural networks. However, simply stacking more layers of the popular Transformer architecture for machine translation results in poor convergence and high computational overhead. Our empirical analysis suggests that convergence is poor due to gradient vanishing caused by the interaction between residual connections and layer normalization. We propose depth-scaled initialization (DS-Init), which decreases parameter variance at the initialization stage, and reduces output variance of residual connections so as to ease gradient back-propagation through normalization layers. To address computational cost, we propose a merged attention sublayer (MAtt) which combines a simplified averagebased self-attention sublayer and the encoderdecoder attention sublayer on the decoder side. Results on WMT and IWSLT translation tasks with five translation directions show that deep Transformers with DS-Init and MAtt can substantially outperform their base counterpart in terms of BLEU (+1.1 BLEU on average for 12-layer models), while matching the decoding speed of the baseline model thanks to the efficiency improvements of MAtt.