Marianna Apidianaki (University of Pennsylvania)
Abdellah Fourtassi (Aix Marseille University)
Sebastian Padó (University of Stuttgart)
Large language models (LLMs) acquire rich world knowledge from the data they are exposed to during training, in a way that appears to parallel how children learn from the language they hear around them. Indeed, since the introduction of these powerful models, there has been a general feeling among researchers in both NLP and cognitive science that a systematic understanding of how these models work and how they use the knowledge they encode, would shed light on the way humans acquire, represent, and process this same knowledge (and vice versa).
Yet, despite the similarities, there are important differences between machines and humans that have prevented a direct translation of insights from the analysis of LLMs to a deeper understanding of human learning. Chief among these differences is that the size of data required to train LLMs far exceeds -- by several orders of magnitude -- the data children need to acquire sophisticated conceptual structures and meanings. Besides, the engineering-driven architectures of LLMs do not appear to have obvious equivalents in children's cognitive apparatus, at least as studied by standard methods in experimental psychology. Finally, children acquire world knowledge not only via exposure to language but also via sensory experience and social interaction.
This edited volume aims to create a forum of exchange and debate between linguists, cognitive scientists and experts in deep learning, NLP and computational linguistics, on the broad topic of learning in humans and machines. Experts from these communities can contribute with empirical and theoretical papers that advance our understanding of this question. Submissions might address the acquisition of different types of linguistic and world knowledge. Additionally, we invite contributions that characterize and address challenges related to the mismatch between humans and LLMs in terms of the size and nature of input data, and the involved learning and processing mechanisms.Topics include, but are not limited to:
Authors are strongly encouraged to submit a short (max 1 page) abstract of their paper by November 10. Abstracts will be sent to the Guest Editors (e-mails below). Minor modifications to the abstract will still be possible until final submission.
Papers should be formatted according to the Computational Linguistics style guidelines: https://cljournal.org/
We accept both long and short papers. Long papers are between 25 and 40 journal pages in length; short papers are between 15 and 25 pages in length.
Papers for this special issue will be submitted through the CL electronic submission system, just like regular papers: https://cljournal.org/submissions.html
Authors of special issue papers will need to select “Special Issue on LLRP” under the Journal Section heading in the CL submission system. Please note that papers submitted to a special issue undergo the same reviewing process as regular papers.
|Deadline for abstract submission
|: November 10, 2023
|Deadline for paper submissions
|: December 17, 2023
|Notification after 1st round of reviewing
|: February 16, 2024
|Revised versions of the papers
|: April 30, 2024
|: June 10, 2024
|Final version of the papers
|: July 1, 2024
All inquiries should be directed to the guest editors of this special issue.
Computational Linguistics is the longest-running flagship journal of the Association for Computational Linguistics. The journal has a high impact factor: 9.3 in 2022 and 7.778 in 2021. Average time to first decision of regular papers and full survey papers (excluding desk rejects) is 34 days for the period January to May 2023, and 47 days for the period January to December 2022.