@inproceedings{tamari-etal-2022-dyna,
title = "{D}yna-b{A}b{I}: unlocking b{A}b{I}{'}s potential with dynamic synthetic benchmarking",
author = "Tamari, Ronen and
Richardson, Kyle and
Kahlon, Noam and
Sar-shalom, Aviad and
Liu, Nelson F. and
Tsarfaty, Reut and
Shahaf, Dafna",
editor = "Nastase, Vivi and
Pavlick, Ellie and
Pilehvar, Mohammad Taher and
Camacho-Collados, Jose and
Raganato, Alessandro",
booktitle = "Proceedings of the 11th Joint Conference on Lexical and Computational Semantics",
month = jul,
year = "2022",
address = "Seattle, Washington",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.starsem-1.9",
doi = "10.18653/v1/2022.starsem-1.9",
pages = "101--122",
abstract = "While neural language models often perform surprisingly well on natural language understanding (NLU) tasks, their strengths and limitations remain poorly understood. Controlled synthetic tasks are thus an increasingly important resource for diagnosing model behavior. In this work we focus on story understanding, a core competency for NLU systems. However, the main synthetic resource for story understanding, the bAbI benchmark, lacks such a systematic mechanism for controllable task generation. We develop Dyna-bAbI, a dynamic framework providing fine-grained control over task generation in bAbI. We demonstrate our ideas by constructing three new tasks requiring compositional generalization, an important evaluation setting absent from the original benchmark. We tested both special-purpose models developed for bAbI as well as state-of-the-art pre-trained methods, and found that while both approaches solve the original tasks (99{\%} accuracy), neither approach succeeded in the compositional generalization setting, indicating the limitations of the original training data. We explored ways to augment the original data, and found that though diversifying training data was far more useful than simply increasing dataset size, it was still insufficient for driving robust compositional generalization (with 70{\%} accuracy for complex compositions). Our results underscore the importance of highly controllable task generators for creating robust NLU systems through a virtuous cycle of model and data development.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tamari-etal-2022-dyna">
<titleInfo>
<title>Dyna-bAbI: unlocking bAbI’s potential with dynamic synthetic benchmarking</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ronen</namePart>
<namePart type="family">Tamari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kyle</namePart>
<namePart type="family">Richardson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Noam</namePart>
<namePart type="family">Kahlon</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aviad</namePart>
<namePart type="family">Sar-shalom</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nelson</namePart>
<namePart type="given">F</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Reut</namePart>
<namePart type="family">Tsarfaty</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dafna</namePart>
<namePart type="family">Shahaf</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2022-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 11th Joint Conference on Lexical and Computational Semantics</title>
</titleInfo>
<name type="personal">
<namePart type="given">Vivi</namePart>
<namePart type="family">Nastase</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ellie</namePart>
<namePart type="family">Pavlick</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohammad</namePart>
<namePart type="given">Taher</namePart>
<namePart type="family">Pilehvar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jose</namePart>
<namePart type="family">Camacho-Collados</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alessandro</namePart>
<namePart type="family">Raganato</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Seattle, Washington</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>While neural language models often perform surprisingly well on natural language understanding (NLU) tasks, their strengths and limitations remain poorly understood. Controlled synthetic tasks are thus an increasingly important resource for diagnosing model behavior. In this work we focus on story understanding, a core competency for NLU systems. However, the main synthetic resource for story understanding, the bAbI benchmark, lacks such a systematic mechanism for controllable task generation. We develop Dyna-bAbI, a dynamic framework providing fine-grained control over task generation in bAbI. We demonstrate our ideas by constructing three new tasks requiring compositional generalization, an important evaluation setting absent from the original benchmark. We tested both special-purpose models developed for bAbI as well as state-of-the-art pre-trained methods, and found that while both approaches solve the original tasks (99% accuracy), neither approach succeeded in the compositional generalization setting, indicating the limitations of the original training data. We explored ways to augment the original data, and found that though diversifying training data was far more useful than simply increasing dataset size, it was still insufficient for driving robust compositional generalization (with 70% accuracy for complex compositions). Our results underscore the importance of highly controllable task generators for creating robust NLU systems through a virtuous cycle of model and data development.</abstract>
<identifier type="citekey">tamari-etal-2022-dyna</identifier>
<identifier type="doi">10.18653/v1/2022.starsem-1.9</identifier>
<location>
<url>https://aclanthology.org/2022.starsem-1.9</url>
</location>
<part>
<date>2022-07</date>
<extent unit="page">
<start>101</start>
<end>122</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Dyna-bAbI: unlocking bAbI’s potential with dynamic synthetic benchmarking
%A Tamari, Ronen
%A Richardson, Kyle
%A Kahlon, Noam
%A Sar-shalom, Aviad
%A Liu, Nelson F.
%A Tsarfaty, Reut
%A Shahaf, Dafna
%Y Nastase, Vivi
%Y Pavlick, Ellie
%Y Pilehvar, Mohammad Taher
%Y Camacho-Collados, Jose
%Y Raganato, Alessandro
%S Proceedings of the 11th Joint Conference on Lexical and Computational Semantics
%D 2022
%8 July
%I Association for Computational Linguistics
%C Seattle, Washington
%F tamari-etal-2022-dyna
%X While neural language models often perform surprisingly well on natural language understanding (NLU) tasks, their strengths and limitations remain poorly understood. Controlled synthetic tasks are thus an increasingly important resource for diagnosing model behavior. In this work we focus on story understanding, a core competency for NLU systems. However, the main synthetic resource for story understanding, the bAbI benchmark, lacks such a systematic mechanism for controllable task generation. We develop Dyna-bAbI, a dynamic framework providing fine-grained control over task generation in bAbI. We demonstrate our ideas by constructing three new tasks requiring compositional generalization, an important evaluation setting absent from the original benchmark. We tested both special-purpose models developed for bAbI as well as state-of-the-art pre-trained methods, and found that while both approaches solve the original tasks (99% accuracy), neither approach succeeded in the compositional generalization setting, indicating the limitations of the original training data. We explored ways to augment the original data, and found that though diversifying training data was far more useful than simply increasing dataset size, it was still insufficient for driving robust compositional generalization (with 70% accuracy for complex compositions). Our results underscore the importance of highly controllable task generators for creating robust NLU systems through a virtuous cycle of model and data development.
%R 10.18653/v1/2022.starsem-1.9
%U https://aclanthology.org/2022.starsem-1.9
%U https://doi.org/10.18653/v1/2022.starsem-1.9
%P 101-122
Markdown (Informal)
[Dyna-bAbI: unlocking bAbI’s potential with dynamic synthetic benchmarking](https://aclanthology.org/2022.starsem-1.9) (Tamari et al., *SEM 2022)
ACL