Shane Steinert-Threlkeld

C. M. Downey, Xuhui Zhou, Leo Z. Liu et al. · cmu

We formulate and test a technique to use Emergent Communication (EC) with a pre-trained multilingual model to improve on modern Unsupervised NMT systems, especially for low-resource languages. It has been argued that the current dominant paradigm in NLP of pre-training on text-only corpora will not yield robust natural language understanding systems, and the need for grounded, goal-oriented, and interactive language learning has been high lighted. In our approach, we embed a multilingual model (mBART, Liu et al., 2020) into an EC image-reference game, in which the model is incentivized to use multilingual generations to accomplish a vision-grounded task. The hypothesis is that this will align multiple languages to a shared task space. We present two variants of EC Fine-Tuning (Steinert-Threlkeld et al., 2022), one of which outperforms a backtranslation-only baseline in all four languages investigated, including the low-resource language Nepali.

C. M. Downey, Terra Blevins, Nora Goldfine et al. · uw

Pre-trained multilingual language models underpin a large portion of modern NLP tools outside of English. A strong baseline for specializing these models for specific languages is Language-Adaptive Pre-Training (LAPT). However, retaining a large cross-lingual vocabulary and embedding matrix comes at considerable excess computational cost during adaptation. In this study, we propose several simple techniques to replace a cross-lingual vocabulary with a compact, language-specific one. Namely, we address strategies for re-initializing the token embedding matrix after vocabulary specialization. We then provide a systematic experimental comparison of our techniques, in addition to the recently-proposed Focus method. We demonstrate that: 1) Embedding-replacement techniques in the monolingual transfer literature are inadequate for adapting multilingual models. 2) Replacing cross-lingual vocabularies with smaller specialized ones provides an efficient method to improve performance in low-resource languages. 3) Simple embedding re-initialization techniques based on script-wise sub-distributions rival techniques such as Focus, which rely on similarity scores obtained from an auxiliary model.

Pangbo Ban, Yifan Jiang, Tianran Liu et al.

To what extent do pre-trained language models grasp semantic knowledge regarding the phenomenon of distributivity? In this paper, we introduce DistNLI, a new diagnostic dataset for natural language inference that targets the semantic difference arising from distributivity, and employ the causal mediation analysis framework to quantify the model behavior and explore the underlying mechanism in this semantically-related task. We find that the extent of models' understanding is associated with model size and vocabulary size. We also provide insights into how models encode such high-level semantic knowledge.

David K. Yi, James V. Bruno, Jiayu Han et al.

We investigate the extent to which verb alternation classes, as described by Levin (1993), are encoded in the embeddings of Large Pre-trained Language Models (PLMs) such as BERT, RoBERTa, ELECTRA, and DeBERTa using selectively constructed diagnostic classifiers for word and sentence-level prediction tasks. We follow and expand upon the experiments of Kann et al. (2019), which aim to probe whether static embeddings encode frame-selectional properties of verbs. At both the word and sentence level, we find that contextual embeddings from PLMs not only outperform non-contextual embeddings, but achieve astonishingly high accuracies on tasks across most alternation classes. Additionally, we find evidence that the middle-to-upper layers of PLMs achieve better performance on average than the lower layers across all probing tasks.

Jingnong Qu, Ashvin Ranjan, Shane Steinert-Threlkeld

Recent breakthroughs in language models (LMs) using neural networks have raised the question: how similar are these models' processing to human language processing? Results using a framework called Brain Score (BS) -- predicting fMRI activations during reading from LM activations -- have been used to argue for a high degree of similarity. To understand this similarity, we conduct experiments by training LMs on various types of input data and evaluate them on BS. We find that models trained on various natural languages from many different language families have very similar BS performance. LMs trained on other structured data -- the human genome, Python, and pure hierarchical structure (nested parentheses) -- also perform reasonably well and close to natural languages in some cases. These findings suggest that BS can highlight language models' ability to extract common structure across natural languages, but that the metric may not be sensitive enough to allow us to infer human-like processing from a high BS score alone.

Abhinav Patil, Jaap Jumelet, Yu Ying Chiu et al. · uw

This paper introduces Filtered Corpus Training, a method that trains language models (LMs) on corpora with certain linguistic constructions filtered out from the training data, and uses it to measure the ability of LMs to perform linguistic generalization on the basis of indirect evidence. We apply the method to both LSTM and Transformer LMs (of roughly comparable size), developing filtered corpora that target a wide range of linguistic phenomena. Our results show that while transformers are better qua LMs (as measured by perplexity), both models perform equally and surprisingly well on linguistic generalization measures, suggesting that they are capable of generalizing from indirect evidence.

Iskar Deng, Nathalia Xu, Shane Steinert-Threlkeld

Recent work has shown that language models (LMs) trained on synthetic corpora can exhibit typological preferences that resemble cross-linguistic regularities in human languages, particularly for syntactic phenomena such as word order. In this paper, we extend this paradigm to differential argument marking (DAM), a semantic licensing system in which morphological marking depends on semantic prominence. Using a controlled synthetic learning method, we train GPT-2 models on 18 corpora implementing distinct DAM systems and evaluate their generalization using minimal pairs. Our results reveal a dissociation between two typological dimensions of DAM. Models reliably exhibit human-like preferences for natural markedness direction, favoring systems in which overt marking targets semantically atypical arguments. In contrast, models do not reproduce the strong object preference in human languages, in which overt marking in DAM more often targets objects rather than subjects. These findings suggest that different typological tendencies may arise from distinct underlying sources.

C. M. Downey, Terra Blevins, Dhwani Serai et al. · uw

The "massively-multilingual" training of multilingual models is known to limit their utility in any one language, and they perform particularly poorly on low-resource languages. However, there is evidence that low-resource languages can benefit from targeted multilinguality, where the model is trained on closely related languages. To test this approach more rigorously, we systematically study best practices for adapting a pre-trained model to a language family. Focusing on the Uralic family as a test case, we adapt XLM-R under various configurations to model 15 languages; we then evaluate the performance of each experimental setting on two downstream tasks and 11 evaluation languages. Our adapted models significantly outperform mono- and multilingual baselines. Furthermore, a regression analysis of hyperparameter effects reveals that adapted vocabulary size is relatively unimportant for low-resource languages, and that low-resource languages can be aggressively up-sampled during training at little detriment to performance in high-resource languages. These results introduce new best practices for performing language adaptation in a targeted setting.

Nicolas Guerin, Shane Steinert-Threlkeld, Emmanuel Chemla

Unsupervised on-the-fly back-translation, in conjunction with multilingual pretraining, is the dominant method for unsupervised neural machine translation. Theoretically, however, the method should not work in general. We therefore conduct controlled experiments with artificial languages to determine what properties of languages make back-translation an effective training method, covering lexical, syntactic, and semantic properties. We find, contrary to popular belief, that (i) parallel word frequency distributions, (ii) partially shared vocabulary, and (iii) similar syntactic structure across languages are not sufficient to explain the success of back-translation. We show however that even crude semantic signal (similar lexical fields across languages) does improve alignment of two languages through back-translation. We conjecture that rich semantic dependencies, parallel across languages, are at the root of the success of unsupervised methods based on back-translation. Overall, the success of unsupervised machine translation was far from being analytically guaranteed. Instead, it is another proof that languages of the world share deep similarities, and we hope to show how to identify which of these similarities can serve the development of unsupervised, cross-linguistic tools.

Leroy Z. Wang, R. Thomas McCoy, Shane Steinert-Threlkeld

What factors contribute to the relative success and corresponding difficulties of in-context learning for Large Language Models (LLMs)? Drawing on insights from the literature on human concept learning, we test LLMs on carefully designed concept learning tasks, and show that task performance highly correlates with the Boolean complexity of the concept. This suggests that in-context learning exhibits a learning bias for simplicity in a way similar to humans.

Ashvin Ranjan, Shane Steinert-Threlkeld

Much recent work has argued that the variation in the languages of the world can be explained from the perspective of efficient communication; in particular, languages can be seen as optimally balancing competing pressures to be simple and to be informative. Focusing on the expression of meaning -- semantic typology -- the present paper asks what factors are responsible for successful explanations in terms of efficient communication. Using the Information Bottleneck (IB) approach to formalizing this trade-off, we first demonstrate and analyze a correlation between optimality in the IB sense and a novel generalization of convexity to this setting. In a second experiment, we manipulate various modeling parameters in the IB framework to determine which factors drive the correlation between convexity and optimality. We find that the convexity of the communicative need distribution plays an especially important role. These results move beyond showing that efficient communication can explain aspects of semantic typology into explanations for why that is the case by identifying which underlying factors are responsible.

Shunjie Wang, Shane Steinert-Threlkeld

Despite the fact that Transformers perform well in NLP tasks, recent studies suggest that self-attention is theoretically limited in learning even some regular and context-free languages. These findings motivated us to think about their implications in modeling natural language, which is hypothesized to be mildly context-sensitive. We test the Transformer's ability to learn mildly context-sensitive languages of varying complexities, and find that they generalize well to unseen in-distribution data, but their ability to extrapolate to longer strings is worse than that of LSTMs. Our analyses show that the learned self-attention patterns and representations modeled dependency relations and demonstrated counting behavior, which may have helped the models solve the languages.

Yifan Jiang, Shane Steinert-Threlkeld

Feature attribution aims to explain the reasoning behind a black-box model's prediction by identifying the impact of each feature on the prediction. Recent work has extended feature attribution to interactions between multiple features. However, the lack of a unified framework has led to a proliferation of methods that are often not directly comparable. This paper introduces a parameterized attribution framework -- the Weighted Möbius Score -- and (i) shows that many different attribution methods for both individual features and feature interactions are special cases and (ii) identifies some new methods. By studying the vector space of attribution methods, our framework utilizes standard linear algebra tools and provides interpretations in various fields, including cooperative game theory and causal mediation analysis. We empirically demonstrate the framework's versatility and effectiveness by applying these attribution methods to feature interactions in sentiment analysis and chain-of-thought prompting.

Jaap Jumelet, Milica Denić, Jakub Szymanik et al.

We investigate the semantic knowledge of language models (LMs), focusing on (1) whether these LMs create categories of linguistic environments based on their semantic monotonicity properties, and (2) whether these categories play a similar role in LMs as in human language understanding, using negative polarity item licensing as a case study. We introduce a series of experiments consisting of probing with diagnostic classifiers (DCs), linguistic acceptability tasks, as well as a novel DC ranking method that tightly connects the probing results to the inner workings of the LM. By applying our experimental pipeline to LMs trained on various filtered corpora, we are able to gain stronger insights into the semantic generalizations that are acquired by these models.

Naomi Tachikawa Shapiro, Amandalynne Paullada, Shane Steinert-Threlkeld

We introduce a multilabel probing task to assess the morphosyntactic representations of word embeddings from multilingual language models. We demonstrate this task with multilingual BERT (Devlin et al., 2018), training probes for seven typologically diverse languages of varying morphological complexity: Afrikaans, Croatian, Finnish, Hebrew, Korean, Spanish, and Turkish. Through this simple but robust paradigm, we show that multilingual BERT renders many morphosyntactic features easily and simultaneously extractable (e.g., gender, grammatical case, pronominal type). We further evaluate the probes on six "held-out" languages in a zero-shot transfer setting: Arabic, Chinese, Marathi, Slovenian, Tagalog, and Yoruba. This style of probing has the added benefit of revealing the linguistic properties that language models recognize as being shared across languages. For instance, the probes performed well on recognizing nouns in the held-out languages, suggesting that multilingual BERT has a conception of noun-hood that transcends individual languages; yet, the same was not true of adjectives.

C. M. Downey, Fei Xia, Gina-Anne Levow et al.

Segmentation remains an important preprocessing step both in languages where "words" or other important syntactic/semantic units (like morphemes) are not clearly delineated by white space, as well as when dealing with continuous speech data, where there is often no meaningful pause between words. Near-perfect supervised methods have been developed for use in resource-rich languages such as Chinese, but many of the world's languages are both morphologically complex, and have no large dataset of "gold" segmentations into meaningful units. To solve this problem, we propose a new type of Segmental Language Model (Sun and Deng, 2018; Kawakami et al., 2019; Wang et al., 2021) for use in both unsupervised and lightly supervised segmentation tasks. We introduce a Masked Segmental Language Model (MSLM) built on a span-masking transformer architecture, harnessing the power of a bi-directional masked modeling context and attention. In a series of experiments, our model consistently outperforms Recurrent SLMs on Chinese (PKU Corpus) in segmentation quality, and performs similarly to the Recurrent model on English (PTB). We conclude by discussing the different challenges posed in segmenting phonemic-type writing systems.

Chih-chan Tien, Shane Steinert-Threlkeld

This work presents methods for learning cross-lingual sentence representations using paired or unpaired bilingual texts. We hypothesize that the cross-lingual alignment strategy is transferable, and therefore a model trained to align only two languages can encode multilingually more aligned representations. We thus introduce dual-pivot transfer: training on one language pair and evaluating on other pairs. To study this theory, we design unsupervised models trained on unpaired sentences and single-pair supervised models trained on bitexts, both based on the unsupervised language model XLM-R with its parameters frozen. The experiments evaluate the models as universal sentence encoders on the task of unsupervised bitext mining on two datasets, where the unsupervised model reaches the state of the art of unsupervised retrieval, and the alternative single-pair supervised model approaches the performance of multilingually supervised models. The results suggest that bilingual training techniques as proposed can be applied to get sentence representations with multilingual alignment.

Chuanrong Li, Lin Shengshuo, Leo Z. Liu et al.

Although large-scale pretrained language models, such as BERT and RoBERTa, have achieved superhuman performance on in-distribution test sets, their performance suffers on out-of-distribution test sets (e.g., on contrast sets). Building contrast sets often re-quires human-expert annotation, which is expensive and hard to create on a large scale. In this work, we propose a Linguistically-Informed Transformation (LIT) method to automatically generate contrast sets, which enables practitioners to explore linguistic phenomena of interests as well as compose different phenomena. Experimenting with our method on SNLI and MNLI shows that current pretrained language models, although being claimed to contain sufficient linguistic knowledge, struggle on our automatically generated contrast sets. Furthermore, we improve models' performance on the contrast sets by apply-ing LIT to augment the training data, without affecting performance on the original data.

Nur Geffen Lan, Emmanuel Chemla, Shane Steinert-Threlkeld

We propose a general framework to study language emergence through signaling games with neural agents. Using a continuous latent space, we are able to (i) train using backpropagation, (ii) show that discrete messages nonetheless naturally emerge. We explore whether categorical perception effects follow and show that the messages are not compositional.

Shane Steinert-Threlkeld

All natural languages exhibit a distinction between content words (like nouns and adjectives) and function words (like determiners, auxiliaries, prepositions). Yet surprisingly little has been said about the emergence of this universal architectural feature of natural languages. Why have human languages evolved to exhibit this division of labor between content and function words? How could such a distinction have emerged in the first place? This paper takes steps towards answering these questions by showing how the distinction can emerge through reinforcement learning in agents playing a signaling game across contexts which contain multiple objects that possess multiple perceptually salient gradable properties.

Lewis O'Sullivan, Shane Steinert-Threlkeld

How are the meanings of linguistic expressions related to their use in concrete cognitive tasks? Visual identification tasks show human speakers can exhibit considerable variation in their understanding, representation and verification of certain quantifiers. This paper initiates an investigation into neural models of these psycho-semantic tasks. We trained two types of network -- a convolutional neural network (CNN) model and a recurrent model of visual attention (RAM) -- on the "most" verification task from \citet{Pietroski2009}, manipulating the visual scene and novel notions of task duration. Our results qualitatively mirror certain features of human performance (such as sensitivity to the ratio of set sizes, indicating a reliance on approximate number) while differing in interesting ways (such as exhibiting a subtly different pattern for the effect of image type). We conclude by discussing the prospects for using neural models as cognitive models of this and other psychosemantic tasks.

Sandro Pezzelle, Shane Steinert-Threlkeld, Raffaela Bernardi et al.

We study the role of linguistic context in predicting quantifiers (`few', `all'). We collect crowdsourced data from human participants and test various models in a local (single-sentence) and a global context (multi-sentence) condition. Models significantly out-perform humans in the former setting and are only slightly better in the latter. While human performance improves with more linguistic context (especially on proportional quantifiers), model performance suffers. Models are very effective in exploiting lexical and morpho-syntactic patterns; humans are better at genuinely understanding the meaning of the (global) context.