Structural Persistence in Language Models: Priming as a Window into Abstract Language Representations

Arabella Sinclair; Jaap Jumelet; Willem Zuidema; Raquel Fernández

doi:10.1162/tacl_a_00504

Structural Persistence in Language Models: Priming as a Window into Abstract Language Representations

Arabella Sinclair, Jaap Jumelet, Willem Zuidema, Raquel Fernández

University of Amsterdam

Research output: Contribution to journal › Article › peer-review

15 Citations (Scopus)

3 Downloads (Pure)

Abstract

We investigate the extent to which modern neural language models are susceptible to structural priming, the phenomenon whereby the structure of a sentence makes the same structure more probable in a follow-up sentence. We explore how priming can be used to study the potential of these models to learn abstract structural information, which is a prerequisite for good performance on tasks that require natural language understanding skills. We introduce a novel metric and release Prime-LM, a large corpus where we control for various linguistic factors that interact with priming strength. We find that Transformer models indeed show evidence of structural priming, but also that the generalizations they learned are to some extent modulated by semantic information. Our experiments also show that the representations acquired by the models may not only encode abstract sequential structure but involve certain level of hierarchical syntactic information. More generally, our study shows that the priming paradigm is a useful, additional tool for gaining insights into the capacities of language models and opens the door to future priming-based investigations that probe the model’s internal states.¹

Original language	English
Pages (from-to)	1031-1050
Number of pages	20
Journal	Transactions of the Association for Computational Linguistics
Volume	10
Early online date	19 Sept 2022
DOIs	https://doi.org/10.1162/tacl_a_00504
Publication status	Published - 19 Sept 2022
Event	SICSA Lecture on Structural Persistence in Language Models: Priming as a Window into Abstract Language Representations - Edinburgh, United Kingdom Duration: 22 Nov 2022 → 22 Nov 2022 https://www.sicsa.ac.uk/events/lecture-on-structural-persistence-in-language-models-priming-as-a-window-into-abstract-language-representations/

Bibliographical note

Acknowledgments
We would like to thank the anonymous reviewers for their extensive and thoughtful feedback and suggestions, which greatly improved our work, as the action editor for his helpful guidance. We would also like to thank members of the ILLC past and present for their useful comments and feedback, specifically, Dieuwke Hupkes, Mario Giulianelli, Sandro Pezzelle, and Ece Takmaz. Arabella Sinclair worked on this project while affiliated with the University of Amsterdam. The project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 819455).

Data Availability Statement

Our code and data can be found at https://github.com/dmg-illc/prime-lm.

Access to Document

10.1162/tacl_a_00504Licence: CC BY

Sinclair_etal_TACL_Structural_Persistance_In_VoR
© 2022 Association for Computational Linguistics. Distributed under a CC-BY 4.0 license. https://creativecommons.org/licenses/by/4.0/
Final published version, 426 KBLicence: CC BY

Cite this

@article{0cb95227b70345018bb8abe75a86ea4a,

title = "Structural Persistence in Language Models: Priming as a Window into Abstract Language Representations",

abstract = "We investigate the extent to which modern neural language models are susceptible to structural priming, the phenomenon whereby the structure of a sentence makes the same structure more probable in a follow-up sentence. We explore how priming can be used to study the potential of these models to learn abstract structural information, which is a prerequisite for good performance on tasks that require natural language understanding skills. We introduce a novel metric and release Prime-LM, a large corpus where we control for various linguistic factors that interact with priming strength. We find that Transformer models indeed show evidence of structural priming, but also that the generalizations they learned are to some extent modulated by semantic information. Our experiments also show that the representations acquired by the models may not only encode abstract sequential structure but involve certain level of hierarchical syntactic information. More generally, our study shows that the priming paradigm is a useful, additional tool for gaining insights into the capacities of language models and opens the door to future priming-based investigations that probe the model{\textquoteright}s internal states.1",

author = "Arabella Sinclair and Jaap Jumelet and Willem Zuidema and Raquel Fern{\'a}ndez",

note = "Acknowledgments We would like to thank the anonymous reviewers for their extensive and thoughtful feedback and suggestions, which greatly improved our work, as the action editor for his helpful guidance. We would also like to thank members of the ILLC past and present for their useful comments and feedback, specifically, Dieuwke Hupkes, Mario Giulianelli, Sandro Pezzelle, and Ece Takmaz. Arabella Sinclair worked on this project while affiliated with the University of Amsterdam. The project has received funding from the European Research Council (ERC) under the European Union{\textquoteright}s Horizon 2020 research and innovation programme (grant agreement No. 819455).; SICSA Lecture on Structural Persistence in Language Models : Priming as a Window into Abstract Language Representations ; Conference date: 22-11-2022 Through 22-11-2022",

year = "2022",

month = sep,

day = "19",

doi = "10.1162/tacl_a_00504",

language = "English",

volume = "10",

pages = "1031--1050",

journal = "Transactions of the Association for Computational Linguistics",

publisher = "The MIT Press",

url = "https://www.sicsa.ac.uk/events/lecture-on-structural-persistence-in-language-models-priming-as-a-window-into-abstract-language-representations/",

}

TY - JOUR

T1 - Structural Persistence in Language Models

T2 - SICSA Lecture on Structural Persistence in Language Models

AU - Sinclair, Arabella

AU - Jumelet, Jaap

AU - Zuidema, Willem

AU - Fernández, Raquel

N1 - Acknowledgments We would like to thank the anonymous reviewers for their extensive and thoughtful feedback and suggestions, which greatly improved our work, as the action editor for his helpful guidance. We would also like to thank members of the ILLC past and present for their useful comments and feedback, specifically, Dieuwke Hupkes, Mario Giulianelli, Sandro Pezzelle, and Ece Takmaz. Arabella Sinclair worked on this project while affiliated with the University of Amsterdam. The project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 819455).

PY - 2022/9/19

Y1 - 2022/9/19

N2 - We investigate the extent to which modern neural language models are susceptible to structural priming, the phenomenon whereby the structure of a sentence makes the same structure more probable in a follow-up sentence. We explore how priming can be used to study the potential of these models to learn abstract structural information, which is a prerequisite for good performance on tasks that require natural language understanding skills. We introduce a novel metric and release Prime-LM, a large corpus where we control for various linguistic factors that interact with priming strength. We find that Transformer models indeed show evidence of structural priming, but also that the generalizations they learned are to some extent modulated by semantic information. Our experiments also show that the representations acquired by the models may not only encode abstract sequential structure but involve certain level of hierarchical syntactic information. More generally, our study shows that the priming paradigm is a useful, additional tool for gaining insights into the capacities of language models and opens the door to future priming-based investigations that probe the model’s internal states.1

AB - We investigate the extent to which modern neural language models are susceptible to structural priming, the phenomenon whereby the structure of a sentence makes the same structure more probable in a follow-up sentence. We explore how priming can be used to study the potential of these models to learn abstract structural information, which is a prerequisite for good performance on tasks that require natural language understanding skills. We introduce a novel metric and release Prime-LM, a large corpus where we control for various linguistic factors that interact with priming strength. We find that Transformer models indeed show evidence of structural priming, but also that the generalizations they learned are to some extent modulated by semantic information. Our experiments also show that the representations acquired by the models may not only encode abstract sequential structure but involve certain level of hierarchical syntactic information. More generally, our study shows that the priming paradigm is a useful, additional tool for gaining insights into the capacities of language models and opens the door to future priming-based investigations that probe the model’s internal states.1

UR - https://aclanthology.org/2022.tacl-1.60/

U2 - 10.1162/tacl_a_00504

DO - 10.1162/tacl_a_00504

M3 - Article

VL - 10

SP - 1031

EP - 1050

JO - Transactions of the Association for Computational Linguistics

JF - Transactions of the Association for Computational Linguistics

Y2 - 22 November 2022 through 22 November 2022

ER -

Structural Persistence in Language Models: Priming as a Window into Abstract Language Representations

Abstract

Bibliographical note

Data Availability Statement

Access to Document

Other files and links

Fingerprint

Cite this