@inproceedings{bommasani-etal-2020-interpreting,
title = "{I}nterpreting {P}retrained {C}ontextualized {R}epresentations via {R}eductions to {S}tatic {E}mbeddings",
author = "Bommasani, Rishi and
Davis, Kelly and
Cardie, Claire",
editor = "Jurafsky, Dan and
Chai, Joyce and
Schluter, Natalie and
Tetreault, Joel",
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
month = jul,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2020.acl-main.431/",
doi = "10.18653/v1/2020.acl-main.431",
pages = "4758--4781",
abstract = "Contextualized representations (e.g. ELMo, BERT) have become the default pretrained representations for downstream NLP applications. In some settings, this transition has rendered their static embedding predecessors (e.g. Word2Vec, GloVe) obsolete. As a side-effect, we observe that older interpretability methods for static embeddings {---} while more diverse and mature than those available for their dynamic counterparts {---} are underutilized in studying newer contextualized representations. Consequently, we introduce simple and fully general methods for converting from contextualized representations to static lookup-table embeddings which we apply to 5 popular pretrained models and 9 sets of pretrained weights. Our analysis of the resulting static embeddings notably reveals that pooling over many contexts significantly improves representational quality under intrinsic evaluation. Complementary to analyzing representational quality, we consider social biases encoded in pretrained representations with respect to gender, race/ethnicity, and religion and find that bias is encoded disparately across pretrained models and internal layers even for models with the same training data. Concerningly, we find dramatic inconsistencies between social bias estimators for word embeddings."
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bommasani-etal-2020-interpreting">
<titleInfo>
<title>Interpreting Pretrained Contextualized Representations via Reductions to Static Embeddings</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rishi</namePart>
<namePart type="family">Bommasani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kelly</namePart>
<namePart type="family">Davis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Claire</namePart>
<namePart type="family">Cardie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2020-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dan</namePart>
<namePart type="family">Jurafsky</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joyce</namePart>
<namePart type="family">Chai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Natalie</namePart>
<namePart type="family">Schluter</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joel</namePart>
<namePart type="family">Tetreault</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Online</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Contextualized representations (e.g. ELMo, BERT) have become the default pretrained representations for downstream NLP applications. In some settings, this transition has rendered their static embedding predecessors (e.g. Word2Vec, GloVe) obsolete. As a side-effect, we observe that older interpretability methods for static embeddings — while more diverse and mature than those available for their dynamic counterparts — are underutilized in studying newer contextualized representations. Consequently, we introduce simple and fully general methods for converting from contextualized representations to static lookup-table embeddings which we apply to 5 popular pretrained models and 9 sets of pretrained weights. Our analysis of the resulting static embeddings notably reveals that pooling over many contexts significantly improves representational quality under intrinsic evaluation. Complementary to analyzing representational quality, we consider social biases encoded in pretrained representations with respect to gender, race/ethnicity, and religion and find that bias is encoded disparately across pretrained models and internal layers even for models with the same training data. Concerningly, we find dramatic inconsistencies between social bias estimators for word embeddings.</abstract>
<identifier type="citekey">bommasani-etal-2020-interpreting</identifier>
<identifier type="doi">10.18653/v1/2020.acl-main.431</identifier>
<location>
<url>https://aclanthology.org/2020.acl-main.431/</url>
</location>
<part>
<date>2020-07</date>
<extent unit="page">
<start>4758</start>
<end>4781</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Interpreting Pretrained Contextualized Representations via Reductions to Static Embeddings
%A Bommasani, Rishi
%A Davis, Kelly
%A Cardie, Claire
%Y Jurafsky, Dan
%Y Chai, Joyce
%Y Schluter, Natalie
%Y Tetreault, Joel
%S Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics
%D 2020
%8 July
%I Association for Computational Linguistics
%C Online
%F bommasani-etal-2020-interpreting
%X Contextualized representations (e.g. ELMo, BERT) have become the default pretrained representations for downstream NLP applications. In some settings, this transition has rendered their static embedding predecessors (e.g. Word2Vec, GloVe) obsolete. As a side-effect, we observe that older interpretability methods for static embeddings — while more diverse and mature than those available for their dynamic counterparts — are underutilized in studying newer contextualized representations. Consequently, we introduce simple and fully general methods for converting from contextualized representations to static lookup-table embeddings which we apply to 5 popular pretrained models and 9 sets of pretrained weights. Our analysis of the resulting static embeddings notably reveals that pooling over many contexts significantly improves representational quality under intrinsic evaluation. Complementary to analyzing representational quality, we consider social biases encoded in pretrained representations with respect to gender, race/ethnicity, and religion and find that bias is encoded disparately across pretrained models and internal layers even for models with the same training data. Concerningly, we find dramatic inconsistencies between social bias estimators for word embeddings.
%R 10.18653/v1/2020.acl-main.431
%U https://aclanthology.org/2020.acl-main.431/
%U https://doi.org/10.18653/v1/2020.acl-main.431
%P 4758-4781
Markdown (Informal)
[Interpreting Pretrained Contextualized Representations via Reductions to Static Embeddings](https://aclanthology.org/2020.acl-main.431/) (Bommasani et al., ACL 2020)
ACL