@inproceedings{waldendorf-etal-2024-contrastive,
title = "Contrastive Decoding Reduces Hallucinations in Large Multilingual Machine Translation Models",
author = "Waldendorf, Jonas and
Haddow, Barry and
Birch, Alexandra",
editor = "Graham, Yvette and
Purver, Matthew",
booktitle = "Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = mar,
year = "2024",
address = "St. Julian{'}s, Malta",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.eacl-long.155",
pages = "2526--2539",
abstract = "In Neural Machine Translation (NMT), models will sometimes generate repetitive or fluent output that is not grounded in the source sentence. This phenomenon is known as hallucination and is a problem even in large-scale multilingual translation models. We propose to use Contrastive Decoding, an algorithm developed to improve generation from unconditional language models, to mitigate hallucinations in NMT. Specifically, we maximise the log-likelihood difference between a model and the same model with reduced contribution from the encoder outputs. Additionally, we propose an alternative implementation of Contrastive Decoding that dynamically weights the difference based on the maximum probability in the output distribution to reduce the effect of CD when the model is confident of its prediction. We evaluate our methods using the Small (418M) and Medium (1.2B) M2M models across 21 low and medium-resource language pairs. Our results show a 14.6 {\mbox{$\pm$}} 0.5 and 11.0 {\mbox{$\pm$}} 0.6 maximal increase in the mean COMET scores for the Small and Medium models on those sentences for which the M2M models initially generate a hallucination., respectively.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="waldendorf-etal-2024-contrastive">
<titleInfo>
<title>Contrastive Decoding Reduces Hallucinations in Large Multilingual Machine Translation Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jonas</namePart>
<namePart type="family">Waldendorf</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Barry</namePart>
<namePart type="family">Haddow</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexandra</namePart>
<namePart type="family">Birch</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-03</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yvette</namePart>
<namePart type="family">Graham</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matthew</namePart>
<namePart type="family">Purver</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">St. Julian’s, Malta</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>In Neural Machine Translation (NMT), models will sometimes generate repetitive or fluent output that is not grounded in the source sentence. This phenomenon is known as hallucination and is a problem even in large-scale multilingual translation models. We propose to use Contrastive Decoding, an algorithm developed to improve generation from unconditional language models, to mitigate hallucinations in NMT. Specifically, we maximise the log-likelihood difference between a model and the same model with reduced contribution from the encoder outputs. Additionally, we propose an alternative implementation of Contrastive Decoding that dynamically weights the difference based on the maximum probability in the output distribution to reduce the effect of CD when the model is confident of its prediction. We evaluate our methods using the Small (418M) and Medium (1.2B) M2M models across 21 low and medium-resource language pairs. Our results show a 14.6 \pm 0.5 and 11.0 \pm 0.6 maximal increase in the mean COMET scores for the Small and Medium models on those sentences for which the M2M models initially generate a hallucination., respectively.</abstract>
<identifier type="citekey">waldendorf-etal-2024-contrastive</identifier>
<location>
<url>https://aclanthology.org/2024.eacl-long.155</url>
</location>
<part>
<date>2024-03</date>
<extent unit="page">
<start>2526</start>
<end>2539</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Contrastive Decoding Reduces Hallucinations in Large Multilingual Machine Translation Models
%A Waldendorf, Jonas
%A Haddow, Barry
%A Birch, Alexandra
%Y Graham, Yvette
%Y Purver, Matthew
%S Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2024
%8 March
%I Association for Computational Linguistics
%C St. Julian’s, Malta
%F waldendorf-etal-2024-contrastive
%X In Neural Machine Translation (NMT), models will sometimes generate repetitive or fluent output that is not grounded in the source sentence. This phenomenon is known as hallucination and is a problem even in large-scale multilingual translation models. We propose to use Contrastive Decoding, an algorithm developed to improve generation from unconditional language models, to mitigate hallucinations in NMT. Specifically, we maximise the log-likelihood difference between a model and the same model with reduced contribution from the encoder outputs. Additionally, we propose an alternative implementation of Contrastive Decoding that dynamically weights the difference based on the maximum probability in the output distribution to reduce the effect of CD when the model is confident of its prediction. We evaluate our methods using the Small (418M) and Medium (1.2B) M2M models across 21 low and medium-resource language pairs. Our results show a 14.6 \pm 0.5 and 11.0 \pm 0.6 maximal increase in the mean COMET scores for the Small and Medium models on those sentences for which the M2M models initially generate a hallucination., respectively.
%U https://aclanthology.org/2024.eacl-long.155
%P 2526-2539
Markdown (Informal)
[Contrastive Decoding Reduces Hallucinations in Large Multilingual Machine Translation Models](https://aclanthology.org/2024.eacl-long.155) (Waldendorf et al., EACL 2024)
ACL