@inproceedings{bhattacharya-bojar-2023-unveiling,
title = "Unveiling Multilinguality in Transformer Models: Exploring Language Specificity in Feed-Forward Networks",
author = "Bhattacharya, Sunit and
Bojar, Ond{\v{r}}ej",
editor = "Belinkov, Yonatan and
Hao, Sophie and
Jumelet, Jaap and
Kim, Najoung and
McCarthy, Arya and
Mohebbi, Hosein",
booktitle = "Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.blackboxnlp-1.9",
doi = "10.18653/v1/2023.blackboxnlp-1.9",
pages = "120--126",
abstract = "Recent research suggests that the feed-forward module within Transformers can be viewed as a collection of key-value memories, where the keys learn to capture specific patterns from the input based on the training examples. The values then combine the output from the {`}memories{'} of the keys to generate predictions about the next token. This leads to an incremental process of prediction that gradually converges towards the final token choice near the output layers. This interesting perspective raises questions about how multilingual models might leverage this mechanism. Specifically, for autoregressive models trained on two or more languages, do all neurons (across layers) respond equally to all languages? No! Our hypothesis centers around the notion that during pre-training, certain model parameters learn strong language-specific features, while others learn more language-agnostic (shared across languages) features. To validate this, we conduct experiments utilizing parallel corpora of two languages that the model was initially pre-trained on. Our findings reveal that the layers closest to the network{'}s input or output tend to exhibit more language-specific behaviour compared to the layers in the middle.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bhattacharya-bojar-2023-unveiling">
<titleInfo>
<title>Unveiling Multilinguality in Transformer Models: Exploring Language Specificity in Feed-Forward Networks</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sunit</namePart>
<namePart type="family">Bhattacharya</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ondřej</namePart>
<namePart type="family">Bojar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2023-12</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yonatan</namePart>
<namePart type="family">Belinkov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sophie</namePart>
<namePart type="family">Hao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jaap</namePart>
<namePart type="family">Jumelet</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Najoung</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arya</namePart>
<namePart type="family">McCarthy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hosein</namePart>
<namePart type="family">Mohebbi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Singapore</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Recent research suggests that the feed-forward module within Transformers can be viewed as a collection of key-value memories, where the keys learn to capture specific patterns from the input based on the training examples. The values then combine the output from the ‘memories’ of the keys to generate predictions about the next token. This leads to an incremental process of prediction that gradually converges towards the final token choice near the output layers. This interesting perspective raises questions about how multilingual models might leverage this mechanism. Specifically, for autoregressive models trained on two or more languages, do all neurons (across layers) respond equally to all languages? No! Our hypothesis centers around the notion that during pre-training, certain model parameters learn strong language-specific features, while others learn more language-agnostic (shared across languages) features. To validate this, we conduct experiments utilizing parallel corpora of two languages that the model was initially pre-trained on. Our findings reveal that the layers closest to the network’s input or output tend to exhibit more language-specific behaviour compared to the layers in the middle.</abstract>
<identifier type="citekey">bhattacharya-bojar-2023-unveiling</identifier>
<identifier type="doi">10.18653/v1/2023.blackboxnlp-1.9</identifier>
<location>
<url>https://aclanthology.org/2023.blackboxnlp-1.9</url>
</location>
<part>
<date>2023-12</date>
<extent unit="page">
<start>120</start>
<end>126</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Unveiling Multilinguality in Transformer Models: Exploring Language Specificity in Feed-Forward Networks
%A Bhattacharya, Sunit
%A Bojar, Ondřej
%Y Belinkov, Yonatan
%Y Hao, Sophie
%Y Jumelet, Jaap
%Y Kim, Najoung
%Y McCarthy, Arya
%Y Mohebbi, Hosein
%S Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP
%D 2023
%8 December
%I Association for Computational Linguistics
%C Singapore
%F bhattacharya-bojar-2023-unveiling
%X Recent research suggests that the feed-forward module within Transformers can be viewed as a collection of key-value memories, where the keys learn to capture specific patterns from the input based on the training examples. The values then combine the output from the ‘memories’ of the keys to generate predictions about the next token. This leads to an incremental process of prediction that gradually converges towards the final token choice near the output layers. This interesting perspective raises questions about how multilingual models might leverage this mechanism. Specifically, for autoregressive models trained on two or more languages, do all neurons (across layers) respond equally to all languages? No! Our hypothesis centers around the notion that during pre-training, certain model parameters learn strong language-specific features, while others learn more language-agnostic (shared across languages) features. To validate this, we conduct experiments utilizing parallel corpora of two languages that the model was initially pre-trained on. Our findings reveal that the layers closest to the network’s input or output tend to exhibit more language-specific behaviour compared to the layers in the middle.
%R 10.18653/v1/2023.blackboxnlp-1.9
%U https://aclanthology.org/2023.blackboxnlp-1.9
%U https://doi.org/10.18653/v1/2023.blackboxnlp-1.9
%P 120-126
Markdown (Informal)
[Unveiling Multilinguality in Transformer Models: Exploring Language Specificity in Feed-Forward Networks](https://aclanthology.org/2023.blackboxnlp-1.9) (Bhattacharya & Bojar, BlackboxNLP-WS 2023)
ACL