@inproceedings{winter-etal-2022-kimera,
title = "{KIMERA}: Injecting Domain Knowledge into Vacant Transformer Heads",
author = {Winter, Benjamin and
Rosero, Alexei Figueroa and
L{\"o}ser, Alexander and
Gers, Felix Alexander and
Siu, Amy},
editor = "Calzolari, Nicoletta and
B{\'e}chet, Fr{\'e}d{\'e}ric and
Blache, Philippe and
Choukri, Khalid and
Cieri, Christopher and
Declerck, Thierry and
Goggi, Sara and
Isahara, Hitoshi and
Maegaard, Bente and
Mariani, Joseph and
Mazo, H{\'e}l{\`e}ne and
Odijk, Jan and
Piperidis, Stelios",
booktitle = "Proceedings of the Thirteenth Language Resources and Evaluation Conference",
month = jun,
year = "2022",
address = "Marseille, France",
publisher = "European Language Resources Association",
url = "https://aclanthology.org/2022.lrec-1.38",
pages = "363--373",
abstract = "Training transformer language models requires vast amounts of text and computational resources. This drastically limits the usage of these models in niche domains for which they are not optimized, or where domain-specific training data is scarce. We focus here on the clinical domain because of its limited access to training data in common tasks, while structured ontological data is often readily available. Recent observations in model compression of transformer models show optimization potential in improving the representation capacity of attention heads. We propose KIMERA (Knowledge Injection via Mask Enforced Retraining of Attention) for detecting, retraining and instilling attention heads with complementary structured domain knowledge. Our novel multi-task training scheme effectively identifies and targets individual attention heads that are least useful for a given downstream task and optimizes their representation with information from structured data. KIMERA generalizes well, thereby building the basis for an efficient fine-tuning. KIMERA achieves significant performance boosts on seven datasets in the medical domain in Information Retrieval and Clinical Outcome Prediction settings. We apply KIMERA to BERT-base to evaluate the extent of the domain transfer and also improve on the already strong results of BioBERT in the clinical domain.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="winter-etal-2022-kimera">
<titleInfo>
<title>KIMERA: Injecting Domain Knowledge into Vacant Transformer Heads</title>
</titleInfo>
<name type="personal">
<namePart type="given">Benjamin</namePart>
<namePart type="family">Winter</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexei</namePart>
<namePart type="given">Figueroa</namePart>
<namePart type="family">Rosero</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Löser</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Felix</namePart>
<namePart type="given">Alexander</namePart>
<namePart type="family">Gers</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amy</namePart>
<namePart type="family">Siu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2022-06</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Thirteenth Language Resources and Evaluation Conference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nicoletta</namePart>
<namePart type="family">Calzolari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Frédéric</namePart>
<namePart type="family">Béchet</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Philippe</namePart>
<namePart type="family">Blache</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Khalid</namePart>
<namePart type="family">Choukri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christopher</namePart>
<namePart type="family">Cieri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thierry</namePart>
<namePart type="family">Declerck</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sara</namePart>
<namePart type="family">Goggi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hitoshi</namePart>
<namePart type="family">Isahara</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bente</namePart>
<namePart type="family">Maegaard</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joseph</namePart>
<namePart type="family">Mariani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hélène</namePart>
<namePart type="family">Mazo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jan</namePart>
<namePart type="family">Odijk</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Stelios</namePart>
<namePart type="family">Piperidis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>European Language Resources Association</publisher>
<place>
<placeTerm type="text">Marseille, France</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Training transformer language models requires vast amounts of text and computational resources. This drastically limits the usage of these models in niche domains for which they are not optimized, or where domain-specific training data is scarce. We focus here on the clinical domain because of its limited access to training data in common tasks, while structured ontological data is often readily available. Recent observations in model compression of transformer models show optimization potential in improving the representation capacity of attention heads. We propose KIMERA (Knowledge Injection via Mask Enforced Retraining of Attention) for detecting, retraining and instilling attention heads with complementary structured domain knowledge. Our novel multi-task training scheme effectively identifies and targets individual attention heads that are least useful for a given downstream task and optimizes their representation with information from structured data. KIMERA generalizes well, thereby building the basis for an efficient fine-tuning. KIMERA achieves significant performance boosts on seven datasets in the medical domain in Information Retrieval and Clinical Outcome Prediction settings. We apply KIMERA to BERT-base to evaluate the extent of the domain transfer and also improve on the already strong results of BioBERT in the clinical domain.</abstract>
<identifier type="citekey">winter-etal-2022-kimera</identifier>
<location>
<url>https://aclanthology.org/2022.lrec-1.38</url>
</location>
<part>
<date>2022-06</date>
<extent unit="page">
<start>363</start>
<end>373</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T KIMERA: Injecting Domain Knowledge into Vacant Transformer Heads
%A Winter, Benjamin
%A Rosero, Alexei Figueroa
%A Löser, Alexander
%A Gers, Felix Alexander
%A Siu, Amy
%Y Calzolari, Nicoletta
%Y Béchet, Frédéric
%Y Blache, Philippe
%Y Choukri, Khalid
%Y Cieri, Christopher
%Y Declerck, Thierry
%Y Goggi, Sara
%Y Isahara, Hitoshi
%Y Maegaard, Bente
%Y Mariani, Joseph
%Y Mazo, Hélène
%Y Odijk, Jan
%Y Piperidis, Stelios
%S Proceedings of the Thirteenth Language Resources and Evaluation Conference
%D 2022
%8 June
%I European Language Resources Association
%C Marseille, France
%F winter-etal-2022-kimera
%X Training transformer language models requires vast amounts of text and computational resources. This drastically limits the usage of these models in niche domains for which they are not optimized, or where domain-specific training data is scarce. We focus here on the clinical domain because of its limited access to training data in common tasks, while structured ontological data is often readily available. Recent observations in model compression of transformer models show optimization potential in improving the representation capacity of attention heads. We propose KIMERA (Knowledge Injection via Mask Enforced Retraining of Attention) for detecting, retraining and instilling attention heads with complementary structured domain knowledge. Our novel multi-task training scheme effectively identifies and targets individual attention heads that are least useful for a given downstream task and optimizes their representation with information from structured data. KIMERA generalizes well, thereby building the basis for an efficient fine-tuning. KIMERA achieves significant performance boosts on seven datasets in the medical domain in Information Retrieval and Clinical Outcome Prediction settings. We apply KIMERA to BERT-base to evaluate the extent of the domain transfer and also improve on the already strong results of BioBERT in the clinical domain.
%U https://aclanthology.org/2022.lrec-1.38
%P 363-373
Markdown (Informal)
[KIMERA: Injecting Domain Knowledge into Vacant Transformer Heads](https://aclanthology.org/2022.lrec-1.38) (Winter et al., LREC 2022)
ACL