The Functional Relevance of Probed Information: A Case Study

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F23%3A10475962" target="_blank" >RIV/00216208:11320/23:10475962 - isvavai.cz</a>

Výsledek na webu

<a href="https://aclanthology.org/2023.eacl-main.58" target="_blank" >https://aclanthology.org/2023.eacl-main.58</a>

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

The Functional Relevance of Probed Information: A Case Study

Popis výsledku v původním jazyce

Recent studies have shown that transformer models like BERT rely on number information encoded in their representations of sentences&apos; subjects and head verbs when performing subject-verb agreement. However, probing experiments suggest that subject number is also encoded in the representations of all words in such sentences. In this paper, we use causal interventions to show that BERT only uses the subject plurality information encoded in its representations of the subject and words that agree with it in number. We also demonstrate that current probing metrics are unable to determine which words&apos; representations contain functionally relevant information. This both provides a revised view of subject-verb agreement in language models, and suggests potential pitfalls for current probe usage and evaluation.

Název v anglickém jazyce

The Functional Relevance of Probed Information: A Case Study

Popis výsledku anglicky

Recent studies have shown that transformer models like BERT rely on number information encoded in their representations of sentences&apos; subjects and head verbs when performing subject-verb agreement. However, probing experiments suggest that subject number is also encoded in the representations of all words in such sentences. In this paper, we use causal interventions to show that BERT only uses the subject plurality information encoded in its representations of the subject and words that agree with it in number. We also demonstrate that current probing metrics are unable to determine which words&apos; representations contain functionally relevant information. This both provides a revised view of subject-verb agreement in language models, and suggests potential pitfalls for current probe usage and evaluation.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

<a href="/cs/project/GA23-06912S" target="_blank" >GA23-06912S: Identifikace a prevence nechtěné genderové zaujatosti v neuronových jazykových modelech</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2023

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics

ISBN

978-1-959429-44-9

ISSN

—

e-ISSN

—

Počet stran výsledku

14

Strana od-do

835-848

Název nakladatele

Association for Computational Linguistics

Místo vydání

Stroudsburg, PA, USA

Místo konání akce

Dubrovnik, Croatia

Datum konání akce

2. 5. 2023

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—