Recent studies have shown that language models pretrained and/or fine-tuned on randomly permuted sentences exhibit competitive performance on GLUE, putting into question the importance of word order information. Somewhat counter-intuitively, some of these studies also report that position embeddings appear to be crucial for models’ good performance with shuffled text. We probe these language models for word order information and investigate what position embeddings learned from shuffled text encode, showing that these models retain a notion of word order information. We show this is in part due to a subtlety in how shuffling is implemented in previous work – before rather than after subword segmentation. Surprisingly, we find even Language models trained on text shuffled after subword segmentation retain some semblance of information about word order because of the statistical dependencies between sentence length and unigram probabilities. Finally, we show that beyond GLUE, a variety of language understanding tasks do require word order information, often to an extent that cannot be learned through fine-tuning.

In an effort to study the inductive biases of language models, numerous studies have attempted to use linguistically motivated tasks as a proxy of sorts, wherein performance on these tasks would imply an inductive bias towards a specific linguistic phenomenon. In this study, we attempt to analyse the inductive biases of language models with respect to natural language phenomena, in the context of building multilingual embedding spaces.We sample corpora from 2 sources in 15 languages and train language models on pseudo-bilingual variants of each corpus, created by duplicating each corpus and shifting token indices for half the resulting corpus. We evaluate the cross-lingual capabilities of these LMs, and show that while correlations with language families tend to be weak, other corpus-level characteristics, such as type-token ratio, tend to be more strongly correlated. Finally, we show that multilingual spaces can be built, albeit less effectively, even when additional destructive perturbations are applied to the training corpora, implying that (effectively) bag-of-words models also have an inductive bias that is sufficient for inducing multilingual spaces.

We analyze the learning dynamics of neural language and translation models using Loss Change Allocation (LCA), an indicator that enables a fine-grained analysis of parameter updates when optimizing for the loss function. In other words, we can observe the contributions of different network components at training time. In this article, we systematically study masked language modeling, causal language modeling, and machine translation. We show that the choice of training objective leads to distinctive optimization procedures, even when performed on comparable Transformer architectures. We demonstrate how the various Transformer parameters are used during training, supporting that the feed-forward components of each layer are the main contributors to the optimization procedure. Finally, we find that the learning dynamics are not affected by data size and distribution but rather determined by the learning objective.

Multilingual pretrained language models are rapidly gaining popularity in NLP systems for non-English languages. Most of these models feature an important corpus sampling step in the process of accumulating training data in different languages, to ensure that the signal from better resourced languages does not drown out poorly resourced ones. In this study, we train multiple multilingual recurrent language models, based on the ELMo architecture, and analyse both the effect of varying corpus size ratios on downstream performance, as well as the performance difference between monolingual models for each language, and broader multilingual language models. As part of this effort, we also make these trained models available for public use.

Since the popularization of the Transformer as a general-purpose feature encoder for NLP, many studies have attempted to decode linguistic structure from its novel multi-head attention mechanism. However, much of such work focused almost exclusively on English — a language with rigid word order and a lack of inflectional morphology. In this study, we present decoding experiments for multilingual BERT across 18 languages in order to test the generalizability of the claim that dependency syntax is reflected in attention patterns. We show that full trees can be decoded above baseline accuracy from single attention heads, and that individual relations are often tracked by the same heads across languages. Furthermore, in an attempt to address recent debates about the status of attention as an explanatory mechanism, we experiment with fine-tuning mBERT on a supervised parsing objective while freezing different series of parameters. Interestingly, in steering the objective to learn explicit linguistic structure, we find much of the same structure represented in the resulting attention patterns, with interesting differences with respect to which parameters are frozen.

In order to preserve word-order information in a non-autoregressive setting, transformer architectures tend to include positional knowledge, by (for instance) adding positional encodings to token embeddings. Several modifications have been proposed over the sinusoidal positional encodings used in the original transformer architecture; these include, for instance, separating position encodings and token embeddings, or directly modifying attention weights based on the distance between word pairs. We first show that surprisingly, while these modifications tend to improve monolingual language models, none of them result in better multilingual language models. We then answer why that is: sinusoidal encodings were explicitly designed to facilitate compositionality by allowing linear projections over arbitrary time steps. Higher variances in multilingual training distributions requires higher compression, in which case, compositionality becomes indispensable. Learned absolute positional encodings (e.g., in mBERT) tend to approximate sinusoidal embeddings in multilingual settings, but more complex positional encoding architectures lack the inductive bias to effectively learn cross-lingual alignment. In other words, while sinusoidal positional encodings were designed for monolingual applications, they are particularly useful in multilingual language models.

Recent work on the interpretability of deep neural language models has concluded that many properties of natural language syntax are encoded in their representational spaces. However, such studies often suffer from limited scope by focusing on a single language and a single linguistic formalism. In this study, we aim to investigate the extent to which the semblance of syntactic structure captured by language models adheres to a surface-syntactic or deep syntactic style of analysis, and whether the patterns are consistent across different languages. We apply a probe for extracting directed dependency trees to BERT and ELMo models trained on 13 different languages, probing for two different syntactic annotation styles: Universal Dependencies (UD), prioritizing deep syntactic relations, and Surface-Syntactic Universal Dependencies (SUD), focusing on surface structure. We find that both models exhibit a preference for UD over SUD — with interesting variations across languages and layers — and that the strength of this preference is correlated with differences in tree shape.

Large-scale pretrained language models are the major driving force behind recent improvements in perfromance on the Winograd Schema Challenge, a widely employed test of commonsense reasoning ability. We show, however, with a new diagnostic dataset, that these models are sensitive to linguistic perturbations of the Winograd examples that minimally affect human understanding. Our results highlight interesting differences between humans and language models: language models are more sensitive to number or gender alternations and synonym replacements than humans, and humans are more stable and consistent in their predictions, maintain a much higher absolute performance, and perform better on non-associative instances than associative ones.

Massively multilingual transformers (MMTs) pretrained via language modeling (e.g., mBERT, XLM-R) have become a default paradigm for zero-shot language transfer in NLP, offering unmatched transfer performance. Current evaluations, however, verify their efficacy in transfers (a) to languages with sufficiently large pretraining corpora, and (b) between close languages. In this work, we analyze the limitations of downstream language transfer with MMTs, showing that, much like cross-lingual word embeddings, they are substantially less effective in resource-lean scenarios and for distant languages. Our experiments, encompassing three lower-level tasks (POS tagging, dependency parsing, NER) and two high-level tasks (NLI, QA), empirically correlate transfer performance with linguistic proximity between source and target languages, but also with the size of target language corpora used in MMT pretraining. Most importantly, we demonstrate that the inexpensive few-shot transfer (i.e., additional fine-tuning on a few target-language instances) is surprisingly effective across the board, warranting more research efforts reaching beyond the limiting zero-shot conditions.

This paper extends the task of probing sentence representations for linguistic insight in a multilingual domain. In doing so, we make two contributions: first, we provide datasets for multilingual probing, derived from Wikipedia, in five languages, viz. English, French, German, Spanish and Russian. Second, we evaluate six sentence encoders for each language, each trained by mapping sentence representations to English sentence representations, using sentences in a parallel corpus. We discover that cross-lingually mapped representations are often better at retaining certain linguistic information than representations derived from English encoders trained on natural language inference (NLI) as a downstream task.

Encoders that generate representations based on context have, in recent years, benefited from adaptations that allow for pre-training on large text corpora. Earlier work on evaluating fixed-length sentence representations has included the use of ‘probing’ tasks, that use diagnostic classifiers to attempt to quantify the extent to which these encoders capture specific linguistic phenomena. The principle of probing has also resulted in extended evaluations that include relatively newer word-level pre-trained encoders. We build on probing tasks established in the literature and comprehensively evaluate and analyse – from a typological perspective amongst others – multilingual variants of existing encoders on probing datasets constructed for 6 non-English languages. Specifically, we probe each layer of a multiple monolingual RNN-based ELMo models, the transformer-based BERT’s cased and uncased multilingual variants, and a variant of BERT that uses a cross-lingual modelling scheme (XLM).

This paper describes the development of the first syntactically-annotated corpus of Breton. The corpus is part of the Universal Dependencies project. In the paper we describe how the corpus was prepared, some Breton-specific constructions that required special treatment, and in addition we give results for parsing Breton using a number of off-the-shelf data-driven parsers.

Progress in the quality of machine translation output calls for new automatic evaluation procedures and metrics. In this paper, we extend the Morpheval protocol introduced by Burlot and Yvon (2017) for the English-to-Czech and English-to-Latvian translation directions to three additional language pairs, and report its use to analyze the results of WMT 2018’s participants for these language pairs. Considering additional, typologically varied source and target languages also enables us to draw some generalizations regarding this morphology-oriented evaluation procedure.

We investigate the effects of multi-task learning using the recently introduced task of semantic tagging. We employ semantic tagging as an auxiliary task for three different NLP tasks: part-of-speech tagging, Universal Dependency parsing, and Natural Language Inference. We compare full neural network sharing, partial neural network sharing, and what we term the learning what to share setting where negative transfer between tasks is less likely. Our findings show considerable improvements for all tasks, particularly in the learning what to share setting which shows consistent gains across all tasks.

We participated to the SemEval-2018 shared task on capturing discriminative attributes (Task 10) with a simple system that ranked 8th amongst the 26 teams that took part in the evaluation. Our final score was 0.67, which is competitive with the winning score of 0.75, particularly given that our system is a zero-shot system that requires no training and minimal parameter optimisation. In addition to describing the submitted system, and discussing the implications of the relative success of such a system on this task, we also report on other, more complex models we experimented with.