Mission: Impossible Language Models

Chomsky and others have very directly claimed that large language models (LLMs) are equally capable of learning languages that are possible and impossible for humans to learn. However, there is very little published experimental evidence to support such a claim. Here, we develop a set of synthetic impossible languages of differing complexity, each designed by systematically altering English data with unnatural word orders and grammar rules. These languages lie on an impossibility continuum: at one end are languages that are inherently impossible, such as random and irreversible shuffles of English words, and on the other, languages that may not be intuitively impossible but are often considered so in linguistics, particularly those with rules based on counting word positions. We report on a wide range of evaluations to assess the capacity of GPT-2 small models to learn these uncontroversially impossible languages, and crucially, we perform these assessments at various stages throughout training to compare the learning process for each language. Our core finding is that GPT-2 struggles to learn impossible languages when compared to English as a control, challenging the core claim. More importantly, we hope our approach opens up a productive line of inquiry in which different LLM architectures are tested on a variety of impossible languages in an effort to learn more about how LLMs can be used as tools for these cognitive and typological investigations.

Separating the Wheat from the Chaff with BREAD: An open-source benchmark and metrics to detect redundancy in text

Data quality is a problem that perpetually resurfaces throughout the field of NLP, regardless of task, domain, or architecture, and remains especially severe for lower-resource languages. A typical and insidious issue, affecting both training data and model output, is data that is repetitive and dominated by linguistically uninteresting boilerplate, such as price catalogs or computer-generated log files. Though this problem permeates many web-scraped corpora, there has yet to be a benchmark to test against, or a systematic study to find simple metrics that generalize across languages and agree with human judgements of data quality. In the present work, we create and release BREAD, a human-labeled benchmark on repetitive boilerplate vs. plausible linguistic content, spanning 360 languages. We release several baseline CRED (Character REDundancy) scores along with it, and evaluate their effectiveness on BREAD. We hope that the community will use this resource to develop better filtering methods, and that our reference implementations of CRED scores can become standard corpus evaluation tools, driving the development of cleaner language modeling corpora, especially in low-resource languages.

Pretrain on just structure: Understanding linguistic inductive biases using transfer learning

Both humans and transformer language models are able to learn language without explicit structural supervision. What inductive learning biases make this learning possible? In this study, we examine the effect of different inductive learning biases by predisposing language models with structural biases through pretraining on artificial structured data, and then evaluating by fine-tuning on English. Our experimental setup gives us the ability to actively control the inductive bias of language models. With our experiments, we investigate the comparative success of three types of inductive bias: 1) an inductive bias for recursive, hierarchical processing 2) an inductive bias for unrestricted token-token dependencies that can't be modeled by context-free grammars, and 3) an inductive bias for a Zipfian power-law vocabulary distribution. We show that complex token-token interactions form the best inductive biases, and that this is strongest in the non-context-free case. We also show that a Zipfian vocabulary distribution forms a good inductive bias independently from grammatical structure. Our study leverages the capabilities of transformer models to run controlled language learning experiments that are not possible to run in humans, and surfaces hypotheses about the structures that facilitate language learning in both humans and machines.

Multilingual BERT has an accent: Evaluating English influences on fluency in multilingual models

While multilingual language models can improve NLP performance on low-resource languages by leveraging higher-resource languages, they also reduce average performance on all languages (the 'curse of multilinguality'). Here we show another problem with multilingual models: grammatical structures in higher-resource languages bleed into lower-resource languages, a phenomenon we call grammatical structure bias. We show this bias via a novel method for comparing the fluency of multilingual models to the fluency of monolingual Spanish and Greek models: testing their preference for two carefully-chosen variable grammatical structures (optional pronoun-drop in Spanish and optional Subject-Verb ordering in Greek). We find that multilingual BERT is biased toward the English-like setting (explicit pronouns and Subject-Verb-Object ordering) as compared to our monolingual control. With our case studies, we hope to bring to light the fine-grained ways in which dominant languages can affect and bias multilingual performance, and encourage more linguistically-aware fluency evaluation.

When classifying grammatical role, BERT doesn't care about word order... except when it matters

Because meaning can often be inferred from lexical semantics alone, word order is often a redundant cue in natural language. For example, the words chopped, chef, and onion are more likely used to convey "The chef chopped the onion," not "The onion chopped the chef." Recent work has shown large language models to be surprisingly word order invariant, but crucially has largely considered natural prototypical inputs, where compositional meaning mostly matches lexical expectations. To overcome this confound, we probe grammatical role representation in English BERT and GPT-2, on instances where lexical expectations are not sufficient, and word order knowledge is necessary for correct classification. Such non-prototypical instances are naturally occurring English sentences with inanimate subjects or animate objects, or sentences where we systematically swap the arguments to make sentences like "The onion chopped the chef". We find that, while early layer embeddings are largely lexical, word order is in fact crucial in defining the later-layer representations of words in semantically non-prototypical positions. Our experiments isolate the effect of word order on the contextualization process, and highlight how models use context in the uncommon, but critical, instances where it matters.

Oolong: Investigating What Makes Crosslingual Transfer Hard with Controlled Studies

Little is known about what makes cross-lingual transfer hard, since factors like tokenization, morphology, and syntax all change at once between languages. To disentangle the impact of these factors, we propose a set of controlled transfer studies: we systematically transform GLUE tasks to alter different factors one at a time, then measure the resulting drops in a pretrained model's downstream performance. In contrast to prior work suggesting little effect from syntax on knowledge transfer, we find significant impacts from syntactic shifts (3-6% drop), though models quickly adapt with continued pretraining on a small dataset. However, we find that by far the most impactful factor for crosslingual transfer is the challenge of aligning the new embeddings with the existing transformer layers (18% drop), with little additional effect from switching tokenizers (<2% drop) or word morphologies (<2% drop). Moreover, continued pretraining with a small dataset is not very effective at closing this gap - suggesting that new directions are needed for solving this problem.

On the Opportunities and Risks of Foundation Models

AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.

Quality at a Glance: An Audit of Web-Crawled Multilingual Datasets

With the success of large-scale pre-training and multilingual modeling in Natural Language Processing (NLP), recent years have seen a proliferation of large, web-mined text datasets covering hundreds of languages. However, to date there has been no systematic analysis of the quality of these publicly available datasets, or whether the datasets actually contain content in the languages they claim to represent. In this work, we manually audit the quality of 205 language-specific corpora released with five major public datasets (CCAligned, ParaCrawl, WikiMatrix, OSCAR, mC4), and audit the correctness of language codes in a sixth (JW300). We find that lower-resource corpora have systematic issues: at least 15 corpora are completely erroneous, and a significant fraction contains less than 50% sentences of acceptable quality. Similarly, we find 82 corpora that are mislabeled or use nonstandard/ambiguous language codes. We demonstrate that these issues are easy to detect even for non-speakers of the languages in question, and supplement the human judgements with automatic analyses. Inspired by our analysis, we recommend techniques to evaluate and improve multilingual corpora and discuss the risks that come with low-quality data releases.

Deep Subjecthood: Higher-Order Grammatical Features in Multilingual BERT

We investigate how Multilingual BERT (mBERT) encodes grammar by examining how the high-order grammatical feature of morphosyntactic alignment (how different languages define what counts as a "subject") is manifested across the embedding spaces of different languages. To understand if and how morphosyntactic alignment affects contextual embedding spaces, we train classifiers to recover the subjecthood of mBERT embeddings in transitive sentences (which do not contain overt information about morphosyntactic alignment) and then evaluate them zero-shot on intransitive sentences (where subjecthood classification depends on alignment), within and across languages. We find that the resulting classifier distributions reflect the morphosyntactic alignment of their training languages. Our results demonstrate that mBERT representations are influenced by high-level grammatical features that are not manifested in any one input sentence, and that this is robust across languages. Further examining the characteristics that our classifiers rely on, we find that features such as passive voice, animacy and case strongly correlate with classification decisions, suggesting that mBERT does not encode subjecthood purely syntactically, but that subjecthood embedding is continuous and dependent on semantic and discourse factors, as is proposed in much of the functional linguistics literature. Together, these results provide insight into how grammatical features manifest in contextual embedding spaces, at a level of abstraction not covered by previous work.

Pretraining on Non-linguistic Structure as a Tool for Analyzing Learning Bias in Language Models

We propose a novel methodology for analyzing the encoding of grammatical structure in neural language models through transfer learning. We test how a language model can leverage its internal representations to transfer knowledge across languages and symbol systems. We train LSTMs on non-linguistic, structured data and test their performance on human language to assess which kinds of data induce generalizable encodings that LSTMs can use for natural language. We find that models trained on structured data such as music and Java code have internal representations that help in modelling human language, and that, surprisingly, adding minimal amounts of structure to the training data makes a large difference in transfer to natural language. Further experiments on transfer between human languages show that zero-shot performance on a test language is highly correlated with syntactic similarity to the training language, even after removing any vocabulary overlap. This suggests that the internal representations induced from natural languages are typologically coherent: they encode the features and differences outlined in typological studies. Our results provide insights into how neural networks represent linguistic structure, and also about the kinds of structural biases that give learners the ability to model language.