@inproceedings{dou-etal-2022-gpt,
title = "Is {GPT}-3 Text Indistinguishable from Human Text? Scarecrow: A Framework for Scrutinizing Machine Text",
author = "Dou, Yao and
Forbes, Maxwell and
Koncel-Kedziorski, Rik and
Smith, Noah A. and
Choi, Yejin",
editor = "Muresan, Smaranda and
Nakov, Preslav and
Villavicencio, Aline",
booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = may,
year = "2022",
address = "Dublin, Ireland",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.acl-long.501",
doi = "10.18653/v1/2022.acl-long.501",
pages = "7250--7274",
abstract = "Modern neural language models can produce remarkably fluent and grammatical text. So much, in fact, that recent work by Clark et al. (2021) has reported that conventional crowdsourcing can no longer reliably distinguish between machine-authored (GPT-3) and human-authored writing. As errors in machine generations become ever subtler and harder to spot, it poses a new challenge to the research community for robust machine text evaluation. We propose a new framework called Scarecrow for scrutinizing machine text via crowd annotation. To support the broad range of real machine errors that can be identified by laypeople, the ten error categories of Scarecrow{---}such as redundancy, commonsense errors, and incoherence{---}are identified through several rounds of crowd annotation experiments without a predefined ontology. We then use Scarecrow to collect over 41k error spans in human-written and machine-generated paragraphs of English language news text. We isolate factors for detailed analysis, including parameter count, training data, and various decoding-time configurations. Our approach successfully quantifies measurable gaps between human authored text and generations from models of several sizes, including fourteen configurations of GPT-3. In addition, our analysis unveils new insights, with detailed rationales provided by laypeople, e.g., that the commonsense capabilities have been improving with larger models while math capabilities have not, and that the choices of simple decoding hyperparameters can make remarkable differences on the perceived quality of machine text. We release our training material, annotation toolkit and dataset at \url{https://yao-dou.github.io/scarecrow/}.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="dou-etal-2022-gpt">
<titleInfo>
<title>Is GPT-3 Text Indistinguishable from Human Text? Scarecrow: A Framework for Scrutinizing Machine Text</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yao</namePart>
<namePart type="family">Dou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maxwell</namePart>
<namePart type="family">Forbes</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rik</namePart>
<namePart type="family">Koncel-Kedziorski</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Noah</namePart>
<namePart type="given">A</namePart>
<namePart type="family">Smith</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yejin</namePart>
<namePart type="family">Choi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2022-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Smaranda</namePart>
<namePart type="family">Muresan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Preslav</namePart>
<namePart type="family">Nakov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aline</namePart>
<namePart type="family">Villavicencio</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Dublin, Ireland</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Modern neural language models can produce remarkably fluent and grammatical text. So much, in fact, that recent work by Clark et al. (2021) has reported that conventional crowdsourcing can no longer reliably distinguish between machine-authored (GPT-3) and human-authored writing. As errors in machine generations become ever subtler and harder to spot, it poses a new challenge to the research community for robust machine text evaluation. We propose a new framework called Scarecrow for scrutinizing machine text via crowd annotation. To support the broad range of real machine errors that can be identified by laypeople, the ten error categories of Scarecrow—such as redundancy, commonsense errors, and incoherence—are identified through several rounds of crowd annotation experiments without a predefined ontology. We then use Scarecrow to collect over 41k error spans in human-written and machine-generated paragraphs of English language news text. We isolate factors for detailed analysis, including parameter count, training data, and various decoding-time configurations. Our approach successfully quantifies measurable gaps between human authored text and generations from models of several sizes, including fourteen configurations of GPT-3. In addition, our analysis unveils new insights, with detailed rationales provided by laypeople, e.g., that the commonsense capabilities have been improving with larger models while math capabilities have not, and that the choices of simple decoding hyperparameters can make remarkable differences on the perceived quality of machine text. We release our training material, annotation toolkit and dataset at https://yao-dou.github.io/scarecrow/.</abstract>
<identifier type="citekey">dou-etal-2022-gpt</identifier>
<identifier type="doi">10.18653/v1/2022.acl-long.501</identifier>
<location>
<url>https://aclanthology.org/2022.acl-long.501</url>
</location>
<part>
<date>2022-05</date>
<extent unit="page">
<start>7250</start>
<end>7274</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Is GPT-3 Text Indistinguishable from Human Text? Scarecrow: A Framework for Scrutinizing Machine Text
%A Dou, Yao
%A Forbes, Maxwell
%A Koncel-Kedziorski, Rik
%A Smith, Noah A.
%A Choi, Yejin
%Y Muresan, Smaranda
%Y Nakov, Preslav
%Y Villavicencio, Aline
%S Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2022
%8 May
%I Association for Computational Linguistics
%C Dublin, Ireland
%F dou-etal-2022-gpt
%X Modern neural language models can produce remarkably fluent and grammatical text. So much, in fact, that recent work by Clark et al. (2021) has reported that conventional crowdsourcing can no longer reliably distinguish between machine-authored (GPT-3) and human-authored writing. As errors in machine generations become ever subtler and harder to spot, it poses a new challenge to the research community for robust machine text evaluation. We propose a new framework called Scarecrow for scrutinizing machine text via crowd annotation. To support the broad range of real machine errors that can be identified by laypeople, the ten error categories of Scarecrow—such as redundancy, commonsense errors, and incoherence—are identified through several rounds of crowd annotation experiments without a predefined ontology. We then use Scarecrow to collect over 41k error spans in human-written and machine-generated paragraphs of English language news text. We isolate factors for detailed analysis, including parameter count, training data, and various decoding-time configurations. Our approach successfully quantifies measurable gaps between human authored text and generations from models of several sizes, including fourteen configurations of GPT-3. In addition, our analysis unveils new insights, with detailed rationales provided by laypeople, e.g., that the commonsense capabilities have been improving with larger models while math capabilities have not, and that the choices of simple decoding hyperparameters can make remarkable differences on the perceived quality of machine text. We release our training material, annotation toolkit and dataset at https://yao-dou.github.io/scarecrow/.
%R 10.18653/v1/2022.acl-long.501
%U https://aclanthology.org/2022.acl-long.501
%U https://doi.org/10.18653/v1/2022.acl-long.501
%P 7250-7274
Markdown (Informal)
[Is GPT-3 Text Indistinguishable from Human Text? Scarecrow: A Framework for Scrutinizing Machine Text](https://aclanthology.org/2022.acl-long.501) (Dou et al., ACL 2022)
ACL