To process novel sentences, language models (LMs) must generalize compositionally—combine familiar elements in new ways. What aspects of a model’s structure promote compositional generalization? Focusing on transformers, we test the hypothesis, motivated by theoretical and empirical work, that deeper transformers generalize more compositionally. Simply adding layers increases the total number of parameters; to address this confound between depth and size, we construct three classes of models which trade off depth for width such that the total number of parameters is kept constant (41M, 134M and 374M parameters). We pretrain all models as LMs and fine-tune them on tasks that test for compositional generalization. We report three main conclusions: (1) after fine-tuning, deeper models generalize more compositionally than shallower models do, but the benefit of additional layers diminishes rapidly; (2) within each family, deeper models show better language modeling performance, but returns are similarly diminishing; (3) the benefits of depth for compositional generalization cannot be attributed solely to better performance on language modeling. Because model latency is approximately linear in the number of layers, these results lead us to the recommendation that, with a given total parameter budget, transformers can be made shallower than is typical without sacrificing performance.

Recent work has proposed explicitly inducing language-wise modularity in multilingual LMs via sparse fine-tuning (SFT) on per-language subnetworks as a means of better guiding cross-lingual sharing. In this paper, we investigate (1) the degree to which language-wise modularity *naturally* arises within models with no special modularity interventions, and (2) how cross-lingual sharing and interference differ between such models and those with explicit SFT-guided subnetwork modularity. In order to do so, we use XLM-R as our multilingual LM. Moreover, to quantify language specialization and cross-lingual interaction, we use a Training Data Attribution method that estimates the degree to which a model’s predictions are influenced by in-language or cross-language training examples. Our results show that language-specialized subnetworks do naturally arise, and that SFT, rather than always increasing modularity, can decrease language specialization of subnetworks in favor of more cross-lingual sharing.

Grapheme-to-phoneme (G2P) conversion is a task that is inherently related to both written and spoken language. Therefore, our submission to the G2P shared task builds off of mSLAM (Bapna et al., 2022), a 600M parameter encoder model pretrained simultaneously on text from 101 languages and speech from 51 languages. For fine-tuning a G2P model, we combined mSLAM’s text encoder, which uses characters as its input tokens, with an uninitialized single-layer RNN-T decoder (Graves, 2012) whose vocabulary is the set of all 381 phonemes appearing in the shared task data. We took an explicitly multilingual approach to modeling the G2P tasks, fine-tuning and evaluating a single model that covered all the languages in each task, and adding language codes as prefixes to the input strings as a means of specifying the language of each example. Our models perform well in the shared task’s “high” setting (in which they were trained on 1,000 words from each language), though they do poorly in the “low” task setting (training on only 100 words from each language). Our models also perform reasonably in the “mixed” setting (training on 100 words in the target language and 1000 words in a related language), hinting that mSLAM’s multilingual pretraining may be enabling useful cross-lingual sharing.

Large multilingual language models typically share their parameters across all languages, which enables cross-lingual task transfer, but learning can also be hindered when training updates from different languages are in conflict. In this article, we propose novel methods for using language-specific subnetworks, which control cross-lingual parameter sharing, to reduce conflicts and increase positive transfer during fine-tuning. We introduce dynamic subnetworks, which are jointly updated with the model, and we combine our methods with meta-learning, an established, but complementary, technique for improving cross-lingual transfer. Finally, we provide extensive analyses of how each of our methods affects the models.

Data scarcity is a crucial issue for the development of highly multilingual NLP systems. Yet for many under-represented languages (ULs) — languages for which NLP research is particularly far behind in meeting user needs — it is feasible to annotate small amounts of data. Motivated by this, we propose XTREME-UP, a benchmark defined by: its focus on the scarce-data scenario rather than zero-shot; its focus on user-centric tasks — tasks with broad adoption by speakers of high-resource languages; and its focus on under-represented languages where this scarce-data scenario tends to be most realistic. XTREME-UP evaluates the capabilities of language models across 88 under-represented languages over 9 key user-centric technologies including ASR, OCR, MT, and information access tasks that are of general utility. We create new datasets for OCR, autocomplete, semantic parsing, and transliteration, and build on and refine existing datasets for other tasks. XTREME-UP provides methodology for evaluating many modeling scenarios including text only, multi-modal (vision, audio, and text), supervised parameter tuning, and in-context learning. We evaluate commonly used models on the benchmark. We release all code and scripts to train and evaluate models.

We present FRMT, a new dataset and evaluation benchmark for Few-shot Region-aware Machine Translation, a type of style-targeted translation. The dataset consists of professional translations from English into two regional variants each of Portuguese and Mandarin Chinese. Source documents are selected to enable detailed analysis of phenomena of interest, including lexically distinct terms and distractor terms. We explore automatic evaluation metrics for FRMT and validate their correlation with expert human evaluation across both region-matched and mismatched rating scenarios. Finally, we present a number of baseline models for this task, and offer guidelines for how researchers can train, evaluate, and compare their own models. Our dataset and evaluation code are publicly available: https://bit.ly/frmt-task.

Multilingual language models (MLMs) are jointly trained on data from many different languages such that representation of individual languages can benefit from other languages’ data. Impressive performance in zero-shot cross-lingual transfer shows that these models are able to exploit this property. Yet, it remains unclear to what extent, and under which conditions, languages rely on each other’s data. To answer this question, we use TracIn (Pruthi et al., 2020), a training data attribution (TDA) method, to retrieve training samples from multilingual data that are most influential for test predictions in a given language. This allows us to analyse cross-lingual sharing mechanisms of MLMs from a new perspective. While previous work studied cross-lingual sharing at the model parameter level, we present the first approach to study it at the data level. We find that MLMs rely on data from multiple languages during fine-tuning and this reliance increases as fine-tuning progresses. We further find that training samples from other languages can both reinforce and complement the knowledge acquired from data of the test language itself.

Text generation metrics that are not robust to dialect variation make it impossible to tell how well systems perform for many groups of users, and can even penalize systems for producing text in lower-resource dialects. In this paper, we introduce a suite of methods to assess whether metrics are dialect robust. These methods show that state-of-the-art metrics are not dialect robust: they often prioritize dialect similarity over semantics, preferring outputs that are semantically incorrect over outputs that match the semantics of the reference but contain dialect differences. As a step towards dialect-robust metrics for text generation, we propose NANO, which introduces regional and language information to the metric’s pretraining. NANO significantly improves dialect robustness while preserving the correlation between automated metrics and human ratings. It also enables a more ambitious approach to evaluation, dialect awareness, in which system outputs are scored by both semantic match to the reference and appropriateness in any specified dialect.

Current image generation models struggle to reliably produce well-formed visual text. In this paper, we investigate a key contributing factor: popular text-to-image models lack character-level input features, making it much harder to predict a word’s visual makeup as a series of glyphs. To quantify this effect, we conduct a series of experiments comparing character-aware vs. character-blind text encoders. In the text-only domain, we find that character-aware models provide large gains on a novel spelling task (WikiSpell). Applying our learnings to the visual domain, we train a suite of image generation models, and show that character-aware variants outperform their character-blind counterparts across a range of novel text rendering tasks (our DrawText benchmark). Our models set a much higher state-of-the-art on visual spelling, with 30+ point accuracy gains over competitors on rare words, despite training on far fewer examples.

Pipelined NLP systems have largely been superseded by end-to-end neural modeling, yet nearly all commonly used models still require an explicit tokenization step. While recent tokenization approaches based on data-derived subword lexicons are less brittle than manually engineered tokenizers, these techniques are not equally suited to all languages, and the use of any fixed vocabulary may limit a model’s ability to adapt. In this paper, we present Canine, a neural encoder that operates directly on character sequences—without explicit tokenization or vocabulary—and a pre-training strategy that operates either directly on characters or optionally uses subwords as a soft inductive bias. To use its finer-grained input effectively and efficiently, Canine combines downsampling, which reduces the input sequence length, with a deep transformer stack, which encodes context. Canine outperforms a comparable mBert model by 5.7 F1 on TyDi QA, a challenging multilingual benchmark, despite having fewer model parameters.

Pre-trained language models perform well on a variety of linguistic tasks that require symbolic reasoning, raising the question of whether such models implicitly represent abstract symbols and rules. We investigate this question using the case study of BERT’s performance on English subject–verb agreement. Unlike prior work, we train multiple instances of BERT from scratch, allowing us to perform a series of controlled interventions at pre-training time. We show that BERT often generalizes well to subject–verb pairs that never occurred in training, suggesting a degree of rule-governed behavior. We also find, however, that performance is heavily influenced by word frequency, with experiments showing that both the absolute frequency of a verb form, as well as the frequency relative to the alternate inflection, are causally implicated in the predictions BERT makes at inference time. Closer analysis of these frequency effects reveals that BERT’s behavior is consistent with a system that correctly applies the SVA rule in general but struggles to overcome strong training priors and to estimate agreement features (singular vs. plural) on infrequent lexical items.

Machine learning has brought striking advances in multilingual natural language processing capabilities over the past year. For example, the latest techniques have improved the state-of-the-art performance on the XTREME multilingual benchmark by more than 13 points. While a sizeable gap to human-level performance remains, improvements have been easier to achieve in some tasks than in others. This paper analyzes the current state of cross-lingual transfer learning and summarizes some lessons learned. In order to catalyze meaningful progress, we extend XTREME to XTREME-R, which consists of an improved set of ten natural language understanding tasks, including challenging language-agnostic retrieval tasks, and covers 50 typologically diverse languages. In addition, we provide a massively multilingual diagnostic suite and fine-grained multi-dataset evaluation capabilities through an interactive public leaderboard to gain a better understanding of such models.

Confidently making progress on multilingual modeling requires challenging, trustworthy evaluations. We present TyDi QA—a question answering dataset covering 11 typologically diverse languages with 204K question-answer pairs. The languages of TyDi QA are diverse with regard to their typology—the set of linguistic features each language expresses—such that we expect models performing well on this set to generalize across a large number of the world’s languages. We present a quantitative analysis of the data quality and example-level qualitative linguistic analyses of observed language phenomena that would not be found in English-only corpora. To provide a realistic information-seeking task and avoid priming effects, questions are written by people who want to know the answer, but don’t know the answer yet, and the data is collected directly in each language without the use of translation.

State-of-the-art multilingual models depend on vocabularies that cover all of the languages the model will expect to see at inference time, but the standard methods for generating those vocabularies are not ideal for massively multilingual applications. In this work, we introduce a novel procedure for multilingual vocabulary generation that combines the separately trained vocabularies of several automatically derived language clusters, thus balancing the trade-off between cross-lingual subword sharing and language-specific vocabularies. Our experiments show improvements across languages on key multilingual benchmark tasks TyDi QA (+2.9 F1), XNLI (+2.1%), and WikiAnn NER (+2.8 F1) and factor of 8 reduction in out-of-vocabulary rate, all without increasing the size of the model or data.

In this paper, we show that Multilingual BERT (M-BERT), released by Devlin et al. (2018) as a single language model pre-trained from monolingual corpora in 104 languages, is surprisingly good at zero-shot cross-lingual model transfer, in which task-specific annotations in one language are used to fine-tune the model for evaluation in another language. To understand why, we present a large number of probing experiments, showing that transfer is possible even to languages in different scripts, that transfer works best between typologically similar languages, that monolingual corpora can train models for code-switching, and that the model can find translation pairs. From these results, we can conclude that M-BERT does create multilingual representations, but that these representations exhibit systematic deficiencies affecting certain language pairs.

Code-switching, the use of more than one language within a single utterance, is ubiquitous in much of the world, but remains a challenge for NLP largely due to the lack of representative data for training models. In this paper, we present a novel model architecture that is trained exclusively on monolingual resources, but can be applied to unseen code-switched text at inference time. The model accomplishes this by jointly maintaining separate word representations for each of the possible languages, or scripts in the case of transliteration, allowing each to contribute to inferences without forcing the model to commit to a language. Experiments on Hindi-English part-of-speech tagging demonstrate that our approach outperforms standard models when training on monolingual text without transliteration, and testing on code-switched text with alternate scripts.

Compositor attribution, the clustering of pages in a historical printed document by the individual who set the type, is a bibliographic task that relies on analysis of orthographic variation and inspection of visual details of the printed page. In this paper, we introduce a novel unsupervised model that jointly describes the textual and visual features needed to distinguish compositors. Applied to images of Shakespeare’s First Folio, our model predicts attributions that agree with the manual judgements of bibliographers with an accuracy of 87%, even on text that is the output of OCR.