Critic-Driven Decoding for Mitigating Hallucinations in Data-to-text Generation

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F23%3A10475866" target="_blank" >RIV/00216208:11320/23:10475866 - isvavai.cz</a>

Result on the web

<a href="http://dx.doi.org/10.18653/v1/2023.emnlp-main.172" target="_blank" >http://dx.doi.org/10.18653/v1/2023.emnlp-main.172</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.18653/v1/2023.emnlp-main.172" target="_blank" >10.18653/v1/2023.emnlp-main.172</a>

Result language

angličtina

Original language name

Critic-Driven Decoding for Mitigating Hallucinations in Data-to-text Generation

Original language description

Hallucination of text ungrounded in the input is a well-known problem in neural data-to-text generation. Many methods have been proposed to mitigate it, but they typically require altering model architecture or collecting additional data, and thus cannot be easily applied to an existing model. In this paper, we explore a new way to mitigate hallucinations by combining the probabilistic output of a generator language model (LM) with the output of a special &quot;text critic&quot; classifier, which guides the generation by assessing the match between the input data and the text generated so far. Our method does not need any changes to the underlying LM&apos;s architecture or training procedure and can thus be combined with any model and decoding operating on word probabilities. The critic does not need any additional training data, using the base LM&apos;s training data and synthetic negative examples. Our experimental results show that our method improves over the baseline on the WebNLG and OpenDialKG benchmarks

Czech name

—

Czech description

—

Type

D - Article in proceedings

CEP classification

—

OECD FORD branch

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

Project

—

Continuities

R - Projekt Ramcoveho programu EK

Publication year

2023

Confidentiality

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

Article name in the collection

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)

ISBN

979-8-89176-060-8

ISSN

—

e-ISSN

—

Number of pages

10

Pages from-to

2853-2862

Publisher name

Association for Computational Linguistics

Place of publication

Stroudsburg, PA, USA

Event location

Singapore

Event date

Dec 6, 2023

Type of event by nationality

WRD - Celosvětová akce

UT code for WoS article

—