Aditya Yedetore

email: [last]@bu.edu

I am a graduate student in the Linguistics Department at Boston University, where I work with Najoung Kim. My research concerns how patterns of linguistic generalization constrain hypotheses about internal computation: in particular, when observed behavior justifies attributing structured representations and structure-sensitive operations to a learning system. I am especially interested in whether properties such as compositionality, systematicity, and productivity provide evidence about the representational and computational organization underlying language learning. My current work compares generalization in neural networks and humans, combining formal analysis with computational experiments to investigate when similar behavioral capacities reflect shared underlying structure, and when similar performance may arise from importantly different computational organizations.

Previously, I studied Cognitive Science and Computer Science at Johns Hopkins University, where I developed an interest in formal models of language, cognition, and computation, especially in how abstract structure can be represented and learned in both biological and artificial systems.

news

Feb 10, 2026	Invited talk on Classical Computation and Connectionist Models at Harvard’s Language and Cognition.
Apr 11, 2025	Poster presentation on Implicit mechanisms for symbol manipulation in RNNs at NENLP.
Feb 14, 2025	Presented on Classical Computation in Connectionist Models for the ANCOR talk series at Brown university.
Oct 08, 2024	New preprint! With Najoung Kim.
Sep 24, 2024	Presented at Harvard’s Language and Cognition talk series.
May 02, 2024	Invited talk at the MIT CPL on cues to hierarchical generalization in neural networks.
Apr 16, 2024	Spotlight talk & poster at NENLP 2024 on semantic bootstrapping in neural networks.
Jul 10, 2023	Presented at ACL 2023 on hierarchical generalization in neural networks.

selected publications

EMNLP

Semantic Training Signals Promote Hierarchical Syntactic Generalization in Transformers

Aditya Yedetore and Najoung Kim

In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Nov 2024

Abs DOI PDF

Neural networks without hierarchical biases often struggle to learn linguistic rules that come naturally to humans. However, neural networks are trained primarily on form alone, while children acquiring language additionally receive data about meaning. Would neural networks generalize more like humans when trained on both form and meaning? We investigate this by examining if Transformers—neural networks without a hierarchical bias—better achieve hierarchical generalization when trained on both form and meaning compared to when trained on form alone. Our results show that Transformers trained on form and meaning do favor the hierarchical generalization more than those trained on form alone, suggesting that statistical learners without hierarchical biases can leverage semantic training signals to bootstrap hierarchical syntactic generalization.
ACL

How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech

Aditya Yedetore, Tal Linzen, Robert Frank, and 1 more author

In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Jul 2023

Abs DOI arXiv PDF Poster

When acquiring syntax, children consistently choose hierarchical rules over competing non-hierarchical possibilities. Is this preference due to a learning bias for hierarchical structure, or due to more general biases that interact with hierarchical cues in children’s linguistic input? We explore these possibilities by training LSTMs and Transformers - two types of neural networks without a hierarchical bias - on data similar in quantity and content to children’s linguistic input: text from the CHILDES corpus. We then evaluate what these models have learned about English yes/no questions, a phenomenon for which hierarchical structure is crucial. We find that, though they perform well at capturing the surface statistics of child-directed speech (as measured by perplexity), both model types generalize in a way more consistent with an incorrect linear rule than the correct hierarchical rule. These results suggest that human-like generalization from text alone requires stronger biases than the general sequence-processing biases of standard neural network architectures.