Any finite set of training data is consistent with an infinite number of hypothetical algorithms that could have generated it. Studies have shown that when human children learn language, they consistently favor hypotheses based on hierarchical syntactic rules without ever encountering disambiguating examples. A recent line of work has inquired as to whether common neural network architectures share this bias, finding that they do so only under special conditions: when syntactically supervised, when pre-trained on massive corpora, or when trained long past convergence. In this paper, we demonstrate, for the first time, neural network architectures that are able to generalize in human-like fashion without any of the aforementioned requirements: stack-augmented neural networks. We test three base architectures (transformer, simple RNN, LSTM) augmented with two styles of stack: the superposition stack of Joulin & Mikolov (2015) and a nondeterministic generalization of it proposed by DuSell & Chiang (2023). We find that transformers with nondeterministic stacks generalize best out of these architectures on a classical question formation task. We also propose a modification to the stack RNN architecture that improves hierarchical generalization. These results suggest that stack-augmented neural networks may be more accurate models of human language acquisition than standard architectures, serving as useful objects of psycholinguistic study. Our code is publicly available.

Language models (LMs), like other neural networks, often favor shortcut heuristics based on surface-level patterns. Although LMs behave like n-gram models early in training, they must eventually learn hierarchical syntactic representations to correctly apply grammatical rules out-of-distribution (OOD). In this work, we use case studies of English grammar to explore how complex, diverse training data drives models to generalize OOD. We construct a framework that unifies our understanding of random variation with training dynamics, rule selection with memorization, and data diversity with complexity. We show that these factors are nuanced, and that intermediate levels of diversity and complexity lead to inconsistent behavior across random seeds and to unstable training dynamics. Our findings emphasize the critical role of training data in shaping generalization patterns and illuminate how competing model strategies lead to inconsistent generalization outcomes across random seeds.

Linguists have long held that a key aspect of natural language syntax is the recursive organization of language units into constituent structures, and research has suggested that current state-of-the-art language models lack an inherent bias towards this feature. A number of alternative models have been proposed to provide inductive biases towards constituency, including the Tree Transformer, which utilizes a modified attention mechanism to organize tokens into constituents. We investigate Tree Transformers to study whether they utilize meaningful and/or useful constituent structures. We pretrain a large Tree Transformer on language modeling in order to investigate the learned constituent tree representations of sentences, finding little evidence for meaningful structures. Next, we evaluate Tree Transformers with similar transformer models on error detection tasks requiring constituent structure. We find that while the Tree Transformer models may slightly outperform at these tasks, there is little evidence to suggest a meaningful improvement. In general, we conclude that there is little evidence to support Tree Transformer as an effective model of syntactic constituency.

Understanding commonsense knowledge is crucial in the field of Natural Language Processing (NLP). However, the presence of demographic terms in commonsense knowledge poses a potential risk of compromising the performance of NLP models. This study aims to investigate and propose methods for enhancing the performance and effectiveness of a commonsense polarization classifier by mitigating the influence of demographic terms. Three methods are introduced in this paper: (1) hierarchical generalization of demographic terms (2) threshold-based augmentation and (3) integration of hierarchical generalization and threshold-based augmentation methods (IHTA). The first method involves replacing demographic terms with more general ones based on a term hierarchy ontology, aiming to mitigate the influence of specific terms. To address the limited bias-related information, the second method measures the polarization of demographic terms by comparing the changes in the model's predictions when these terms are masked versus unmasked. This method augments commonsense sentences containing terms with high polarization values by replacing their predicates with synonyms generated by ChatGPT. The third method combines the two approaches, starting with threshold-based augmentation followed by hierarchical generalization. The experiments show that the first method increases the accuracy over the baseline by 2.33%, and the second one by 0.96% over standard augmentation methods. The IHTA techniques yielded an 8.82% and 9.96% higher accuracy than threshold-based and standard augmentation methods, respectively.

Natural language is structured hierarchically: words are grouped into phrases or constituents, which can be further grouped to form higher-level phrases up to the full sentence. How well do the neural network models trained on language data learn this phrase structure of human language has been a subject of great interest. A flurry of past work have shown that syntax trees can be recovered from recurrent neural network (RNN) and transformer-based models trained on large-scale language corpora (Tenney et al., 2019, Peters et al., 2018, Lin et al., 2019, Wu et al., 2020). While these studies provide useful evidence of the aforementioned phenomenon, they do not shed light on the architectural choices, training paradigms or dataset characteristics that lead models to learn the phrase structure of language. A useful tool to understand these model and dataset specific properties is through the test for hierarchical generalization, i.e., evaluating the capability of a model to generalize to novel syntactic forms, which were unseen during training. A classic problem to test for hierarchical generalization is question formation, where given a declarative sentence, e.g., My walrus does move the dogs that do wait., the task is to transform it into a question: Does my walrus move the dogs that do wait? The task is accomplished by moving one auxiliary verb to the front. The correct choice to move does in this example (rather than do), is predicted both by a hierarchical rule based on the phrase-structure syntax of the sentence, and by a linear rule that says to move the first auxiliary. Hence, as a test for hierarchical generalization, we can ask, for neural networks trained from scratch on data that is consistent with both hierarchical and linear rules (i.e.,

Accurate syntactic representations are essential for robust generalization in natural language. Recent work has found that pre-training can teach language models to rely on hierarchical syntactic features - as opposed to incorrect linear features - when performing tasks after fine-tuning. We test what aspects of pre-training are important for endowing encoder-decoder Transformers with an inductive bias that favors hierarchical syntactic generalizations. We focus on architectural features (depth, width, and number of parameters), as well as the genre and size of the pre-training corpus, diagnosing inductive biases using two syntactic transformation tasks: question formation and passivization, both in English. We find that the number of parameters alone does not explain hierarchical generalization: model depth plays greater role than model width. We also find that pre-training on simpler language, such as child-directed speech, induces a hierarchical bias using an order-of-magnitude less data than pre-training on more typical datasets based on web text or Wikipedia; this suggests that in cognitively plausible language acquisition settings, neural language models may be more data-efficient than previously thought.

For humans, language production and comprehension is sensitive to the hierarchical structure of sentences. In natural language processing, past work has questioned how effectively neural sequence models like transformers capture this hierarchical structure when generalizing to structurally novel inputs. We show that transformer language models can learn to generalize hierarchically after training for extremely long periods -- far beyond the point when in-domain accuracy has saturated. We call this phenomenon \emph{structural grokking}. On multiple datasets, structural grokking exhibits inverted U-shaped scaling in model depth: intermediate-depth models generalize better than both very deep and very shallow transformers. When analyzing the relationship between model-internal properties and grokking, we find that optimal depth for grokking can be identified using the tree-structuredness metric of \citet{murty2023projections}. Overall, our work provides strong evidence that, with extended training, vanilla transformers discover and use hierarchical structure.

Neural language models (LMs) are arguably less data-efficient than humans -- why does this gap occur? In this study, we hypothesize that this gap stems from the learners' accessibility to modalities other than text, specifically, vision. We conducted two complementary experiments (using noisy, realistic data and a simplified, artificial one) toward the advantage of vision in the syntactic generalization of LMs. Our results showed that vision accelerated a proper linguistic generalization in the simplified, artificial setting, but LMs struggled with the noisy, realistic setting. These mixed results indicate several possibilities, e.g., vision can potentially boost language acquisition, but learners' additional visual/linguistic prior knowledge should be needed to robustly make use of raw images for efficient language acquisition.