The DOI matches its Crossref record; 26 of the 35 references it cites were found in a registry or an index; 9 of the 17 sources its claims cite were checked for support.
StereoSet: Measuring stereotypical bias in pretrained language models
Open accessAbstract
Moin Nadeem, Anna Bethke, Siva Reddy. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.
Overview
AI snapshot
Written from the paper’s full text.
TL;DR
StereoSet introduces a large-scale natural English benchmark of 16,995 Context Association Tests across gender, profession, race, and religion to measure stereotypical bias alongside language modeling ability in pretrained language models. It finds that popular models such as BERT, GPT2, ROBERTA, and XLNET show strong stereotypical biases, and that stronger language modeling correlates with greater bias (Spearman ρ = 0.87).
Key findings
- StereoSet contains 321 target terms and 16,995 test instances (triplets): 8,498 intrasentence CATs and 8,497 intersentence CATs across four domains.
- All pretrained models evaluated have higher language modeling scores (lms) than RANDOMLM but also exhibit more stereotypical behavior than RANDOMLM (ss = 50). In Table 4, GPT2-large reaches lms = 89.6 and ss = 62.7; ROBERTA-base has ss = 58.5.
- Language modeling ability and stereotype score are strongly correlated: Spearman rank correlation ρ = 0.87. The ENSEMBLE model is reported as the most biased model (ss = 62.5 in the text; 62.2 in Table 4).
- SENTIMENTLM achieves a high stereotype score (ss = 60.8) but a lower language modeling score (lms = 65.1), indicating sentiment can help detect stereotypical associations but is insufficient for distinguishing meaningful from meaningless sentences.
Methods
- The paper designs the Context Association Test (CAT) with two formats: intrasentence fill-in-the-blank and intersentence discourse-level association. Each instance includes a target context and three associations: stereotype, anti-stereotype, and meaningless.
- Data were crowdsourced via Amazon Mechanical Turk from workers in the United States. Target terms were curated from Wikidata relations for profession, race, and religion, plus gender terms. Each CAT was written by one worker and validated by five others; CATs were retained when at least three validators agreed on labels, yielding 83% of collected CATs.
- Models evaluated include BERT, ROBERTA, XLNET, and GPT2. Evaluation uses likelihood-based and pseudo-likelihood-based scoring, with metrics: language modeling score (lms), stereotype score (ss), and idealized CAT score (icat). For intersentence tasks, Next Sentence Prediction heads were trained for models lacking them.
Limitations
- The paper states that evaluating full language modeling ability is beyond its scope, that StereoSet has no training set by design, and that neutral associations were discarded because crowdworkers had low inter-annotator agreement and often produced implicitly stereotypical sentences.
- The paper states that achieving ideal StereoSet performance does not guarantee a model is unbiased, that icat is not intended as the sole criterion for model selection, and that further research is needed; it also notes lack of computing resources to train the same model on increasing amounts of corpora.
- Evident from the text: the dataset is US English and collected only from US crowdworkers, so the stereotypes are tailored to the United States; it covers only four domains; validation retains CATs with as few as three of five agreeing validators; and some reported numbers differ between the narrative and tables (e.g., GPT2-large lms/ss and ENSEMBLE ss), which may complicate interpretation.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
34 Own result · 8 Own interpretation · 6 Partly supported · 6 Own method · 5 No basis given · 1 Supported · 1 Supported so far · 1 Sources not found · 1 No full text of the sources · accuracy not yet measured
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“they mainly evaluate bias of pretrained language models on a small set of artificial sentences, even though these models are trained on natural data”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Abstract
What the sources say
- Not checked — no full text of the sourcecites May et al., 2019 source
- Partly supportedcites Kurita et al., 2019 sourceExperimental — accuracy not yet measured
The Introduction and Table 1 describe evaluating BERT bias with a small set of simple template sentences, and the paper discusses an earlier sentence-embedding bias test. But it does not establish that existing bias evaluations mainly use artificial sentences; it also examines bias on a Wikipedia-derived pronoun-resolution dataset. Page numbers are unavailable in our copy. (LLM, conf=0.78)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Third, they focus on single word terms and ignore multiword terms like construction worker.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 1 Introduction
What the sources say
- Not checked — no full text of the sourcecites May et al., 2019 source
- Partly supportedcites Kurita et al., 2019 sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) describes this paper’s bias measure as querying BERT for a particular token and illustrates it with single-word targets and attributes. It does not establish that contextual-embedding bias studies generally ignore multiword terms, or specifically discuss “construction worker.” (LLM, conf=0.78)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“current evaluations focus on measuring bias without considering the language modeling ability of a model, which could lead to misleading trust on a model even if it is a poor language model.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Abstract
What the sources say
- Not checked — no full text of the sourcecites May et al., 2019 source
- Partly supportedcites Kurita et al., 2019 sourceExperimental — accuracy not yet measured
The Introduction and §5 Real World Implications discuss existing bias-evaluation methods and say they can give inconsistent or unclear results for contextualized embeddings. But the paper does not say those evaluations neglect a model’s language-modeling ability, or that they could inspire misplaced trust in a poor language model. Its proposed method uses masked-token probabilities to measure bias, not to assess overall language-modeling quality. (LLM, conf=0.78)
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Lastly, they study bias of a model independent of its language modeling ability which could lead to undeserved trust in a model if it is a poor language model.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 1 Introduction
What the sources say
- Not checked — no full text of the sourcecites May et al., 2019 source
- Partly supportedcites Kurita et al., 2019 sourceExperimental — accuracy not yet measured
The Introduction describes measuring BERT’s bias using masked-token predictions and normalizing those predictions against a prior probability. It does not establish that the bias measure is independent of language-modeling ability or discuss undeserved trust in a poor language model. (LLM, conf=0.84)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“First, the context does not reflect the natural usage of a word.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 1 Introduction
What the sources say
- Not checked — no full text of the sourcecites May et al., 2019 source
- Partly supportedcites Kurita et al., 2019 sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) says the authors create simple template sentences to measure bias, supporting the claim that the study uses constructed contexts. It does not say that those contexts fail to reflect a word’s natural usage; that criticism is not established by this source. (LLM, conf=0.79)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Second, they require stereotypical attribute terms to be predefined (e.g., pleasant and unpleasant terms).”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 1 Introduction
What the sources say
- Not checked — no full text of the sourcecites May et al., 2019 source
- Supportedcites Kurita et al., 2019 source
Section 3.2 says the authors test BERT using specified sets of target and attribute words, including pleasant and unpleasant attributes. The Introduction also says their method requires constructing template sentences containing attribute words of interest. These passages support the claim that the bias tests described in this paper use predefined attribute terms. (LLM, conf=0.88)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“While GPT2-large is the most stereotypical model of all pretrained models (60.1),ROBERTA-base is the least stereotypical model (50.5).”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 7 Results and discussion
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“SENTIMENTLM achieves the highest stereotypical score compared to all pretrained models, indicating that sentiment can indeed be exploited to detect stereotypical associations.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 7 Results and discussion
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“These keywords indicate that target terms in gender and race are associated with physical attributes such as beautiful, feminine, masculine, etc.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.4 Dataset analysis
hedgedcorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“As evident in Table 2, people do not always associate stereotypes with negative associations (e.g., Asians are good at math has positive sentiment).”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.4 Dataset analysis
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Probably not a substantial claim. 17
- It may state two separate claims as one. 14
Every source the paper cites for this claim was checked and supports it.
In the paper
“The seminal works of Bolukbasi et al. (2016) and Caliskan et al. (2017) show that word embeddings such as word2vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014) contain stereotypical biases using diagnostic methods like word analogies and association tests.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 1 Introduction
What the sources say
- Supportedcites Bolukbasi et al. (2016) source
Section 4 (pp. 6–7) compares word2vec and GloVe embeddings, finds gender bias in occupation-word associations across both, and describes using analogies to identify stereotypes. This supports the claim attributed to Bolukbasi et al.; the sentence also cites another work, so it need not establish that Bolukbasi et al. used every diagnostic method listed. (LLM, conf=0.94)
- Supportedcites Caliskan et al. (2017) source
The Results (p. 3) report human-like biases in GloVe using association-based tests. Data and training (p. 9) says the authors repeated the tests with word2vec and found statistically significant effects in every experiment. (LLM, conf=0.94)
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Moreover, Aribandi et al. (2021) has shown that psuedo-likelihood has higher variance than likelihood scoring.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Impact of scoring mechanism
What the sources say
- Partly supportedcites Aribandi et al. (2021) sourceExperimental — accuracy not yet measured
Section 3.1 reports that CrowS-Pairs has far higher variance than StereoSet. Section 2.2 identifies CrowS-Pairs’ pseudo-likelihood as one possible reason, but also cites its smaller dataset and greater number of categories. The paper does not establish that pseudo-likelihood generally has higher variance than likelihood scoring. (LLM, conf=0.86)
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“This filtering results in selecting 83% of the CATs, indicating that there is regularity in stereotypical views among the workers.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.3 CATs validation and human agreement
hedgeddirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Among models, GPT2-large is the best performing language model (88.3) followed by GPT2-medium (85.9).”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 7 Results and discussion
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“This aligns with expectations and indicates that multiple annotators use similar attributes.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.4 Dataset analysis
hedgedcorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We crowdsource Stere-oSet, a dataset for associative contexts in English containing 4 target domains, 321 target terms and 16,995 test instances (triplets).”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 1 Introduction
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“As the model size increases, we see that its language modeling ability (lms) increases, and correspondingly its stereotypical score.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 7 Results and discussion
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our NSP classifier achieves an 94.6% accuracy with ROBERTA-base, a 97.1% accuracy with ROBERTA-large, a 93.4% accuracy with XLNET-base and 94.1% accuracy with XL-NET-large.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Additional information
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our NSP classifier obtains a 92.5% accuracy on GPT2-small, 94.2% on GPT2-medium, and 96.1% on GPT2-large.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Additional information
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our fine-tuned model achieves an 92% test accuracy on the Large Movie Review dataset.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > A.2 Fine-Tuning BERT for Sentiment Analysis
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
In the paper
“Our results are comparable to Devlin et al. (2019), with each model obtaining 93-98% accuracy against the test set of 3.5 million examples.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Additional information
What the sources say
- Not checked — no full text of the sourcecites Devlin et al. (2019) source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In practice, we found that collecting neutral contexts are prone to implicit biases and has low inter-annotator agreement (Section 4).”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 2.2 Formulation
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The model is relatively less biased on race than on others (ss = 61.8).”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Domain-wise bias
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Rests on an assumption: The stereotypes elicited from US crowdworkers are representative of stereotypes in the United States.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Surprisingly, the corpora size does not correlate with either lms or ss.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Impact of scoring mechanism
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“The results exhibit similar trends on the development and test sets.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 7 Results and discussion
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“It is important to note that the classifier also classifies neutral sentiment such as My housekeeper is a Mexican as positive.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > A.2 Fine-Tuning BERT for Sentiment Analysis
plaindirectSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The outlier is Muslim, although it has established stereotypes indicated by the high human agreement (see Table 10).”
§ StereoSet: Measuring stereotypical bias in pretrained language models > Domain-wise bias
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Rests on an assumption: The stereotypes elicited from US crowdworkers are representative of stereotypes in the United States.
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“We restrict ourselves to crowdworkers in USA since stereotypes could change based on the country.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4 Dataset Creation
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim.
- Rests on an assumption: Stereotypes could change based on the country.
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“For each context and its associations, we ask five validators to classify each association into a stereotype, an anti-stereotype or a meaningless association.We only retain CATs where at least three validators agree on the labels.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.3 CATs validation and human agreement
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. It may state two separate claims as one.
In the paper
“Caliskan et al. show that male names are more likely to be associated with career terms than female names where the association is measured using embedding similarity.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 1 Introduction
What the sources say
- Not matchedcites Caliskan et al.
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
hedgedcorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
- Not matched
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“However, its language model performance is worse, which is expected, since sentiment alone isn't sufficient to distinguish meaningful and meaningless sentences.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 7 Results and discussion
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper may give no reason why it holds.
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“We split StereoSet based on the target terms: 25% of the target terms and their instances for the development set and 75% for the test set.We ensure terms in the development set and test set are disjoint.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 5.1 Development and test sets
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Furthermore, we hold that some of these neutral sentences aren't truly neutral; the chess player was bearded may inadvertently conceal stereotypes, since both chess players and bearded men are commonly seen as wise.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > A.4 Collecting Neutral Associations
hedgedconditionalSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“We required all crowdworkers to be in the United States and have a HIT acceptance rate greater than 97%.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 9.1 Curation Rationale
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“We collected this data via Amazon Mechanical Turk (AMT), where each example was written by one crowdworker and validated by four other crowdworkers.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 9.1 Curation Rationale
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“For all models, we train identical Next Sentence Prediction heads on identical datasets (details given in Appendix A.5), and compute the log likelihood that any given target sentence follows the context.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 6.1 Masked Language Models
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim.
The paper relies on this without support.
Stereotypes could change based on the country.
“We restrict ourselves to crowdworkers in USA since stereotypes could change based on the country.”
The authors state this as an assumption.
Language modeling ability and stereotypical bias are of equal importance for model comparison.
“we introduce a simple point-estimate called the idealized CAT (icat) score for model comparison assuming equal importance to language modeling ability and stereotypical bias.”
The paper relies on this without support.
The real world distribution of corpora likely reflects stereotypes.
“The correlation between lms and ss is unfortunate and perhaps un- avoidable as long as we rely on the real world distribution of corpora to train language models since these corpora are likely to reflect stereotypes.”
The paper relies on this without support.
BERT has been trained with a Next Sentence Prediction classification head to provide p(s|c), while the other models have not.
“While BERT has been trained with a Next Sentence Prediction classification head to provide p(s|c), the other models have not.”
The paper relies on this without support.
The stereotypes elicited from US crowdworkers are representative of stereotypes in the United States.
“StereoSet is in English and is tailored for the stereotypes that exist in the United States.”
The paper relies on this without support.
The NSP classifier's accuracy on a Wikipedia-derived corpus means it can measure language modeling ability on StereoSet.
“Our results are comparable to Devlin et al. (2019), with each model obtaining 93-98% accuracy against the test set of 3.5 million examples.”
The paper relies on this without support.
The sentiment classifier's accuracy on movie reviews means it can classify sentiment in StereoSet.
“Our fine-tuned model achieves an 92% test accuracy on the Large Movie Review dataset.”
The paper relies on this without support.
Human agreement scores in Table 10 are a proxy for most and least biased terms.
“see Table 10 for humanagreement scores, a proxy for most and least biased terms”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“For each type of bias, the dataset presents a context sentence with three options: one with a stereotype, one with a neutral or positive connotation (”
From Bias and Fairness in Large Language Models: A Survey (2024) · cites this paper as Nadeem, Bethke, and Reddy 2021
The cited source says what the claim attributes to it.
In the citing paper
“StereoSet presents 16,995 crowdsourced instances measuring race, gender, religion, and profession stereotypes.”
From Bias and Fairness in Large Language Models: A Survey (2024) · cites this paper as Nadeem, Bethke, and Reddy 2021
The cited source says what the claim attributes to it.
In the citing paper
“StereoSet evaluates intrasentence bias within a sentence with fill-in-the-blank sentences, where the options describe a social group in the sentence context”
From Bias and Fairness in Large Language Models: A Survey (2024) · cites this paper as Nadeem, Bethke, and Reddy 2021
The cited source says what the claim attributes to it.
In the citing paper
“It measures intersentence bias between sentences in a discourse with three continuation options, where the first sentence mentions a social group.”
From Bias and Fairness in Large Language Models: A Survey (2024) · cites this paper as Nadeem, Bethke, and Reddy 2021
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
“Intersentence Gender 40 996 15.55 Profession 120 3,269 16.05 Race 149 3,989 14.98 Religion 12 604 14.99 Total 321 8,497 15.39”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.3 CATs validation and human agreement
Reported 8497 · recomputed 8858
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
“Intrasentence Gender 40 1,026 7.98 Profession 120 3,208 8.30 Race 149 3,996 7.63 Religion 12 623 8.18 Total 321 8,498 8.02”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.3 CATs validation and human agreement
Reported 8498 · recomputed 8853
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.3 CATs validation and human agreement
“Religion 12 623 8.18”
§ StereoSet: Measuring stereotypical bias in pretrained language models > A.5 General Methods for Training a Next Sentence Prediction Head
“Religion Jihad (0.86) Muslim (0.84) Holy Trinity (0.81) Quran (0.8) Trinity (0.8) Sanskrit (0.8) Mecca (0.8) Islam (0.79) baptize (0.79) Mohammed (0.79) Sharia (0.78) church (0.77) Jesus (0.77) Christ (0.77) Messiah (0.76) Vishnu (0.76) Hajj (0.76) Bible (0.75) Christian (0.74) Reincarnation (0.74) Hindu (0.74) Brahmin (0.74)Ten Commandments (0.72) Shiva (0.72)”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ StereoSet: Measuring stereotypical bias in pretrained language models > 4.3 CATs validation and human agreement
“For each context and its associations, we ask five validators to classify each association into a stereotype, an anti-stereotype or a meaningless association.”
§ StereoSet: Measuring stereotypical bias in pretrained language models > 9.1 Curation Rationale
“each example was written by one crowdworker and validated by four other crowdworkers.”
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Glove: Global Vectors for Word Representation
Pennington, J., Socher, R., & Manning, C. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1162 · Cited by 34,071
Tversky, A., & Kahneman, D. (1974) · Science · doi:10.1126/science.185.4157.1124 · Cited by 28,810
Distributed Representations of Words and Phrases and their Compositionality
Mikolov, T., Sutskever, I., Chen, K., Corrado, G., & Dean, J. (2013) · arXiv · doi:10.48550/arxiv.1310.4546 · Cited by 24,533
RoBERTa: A Robustly Optimized BERT Pretraining Approach
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019) · arXiv · doi:10.48550/arxiv.1907.11692 · Cited by 16,515
Deep Contextualized Word Representations
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) · doi:10.18653/v1/n18-1202 · Cited by 11,983
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021) · Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3442188.3445922 · Cited by 6,988
XLNet: Generalized Autoregressive Pretraining for Language Understanding
Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R., & Le, Q. V. (2019) · arXiv · doi:10.48550/arxiv.1906.08237 · Cited by 6,687
Implicit social cognition: Attitudes, self-esteem, and stereotypes
Greenwald, A. G., & Banaji, M. R. (1995) · Psychological Review · doi:10.1037/0033-295x.102.1.4 · Cited by 6,682
Universal Language Model Fine-tuning for Text Classification
Howard, J., & Ruder, S. (2018) · Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p18-1031 · Cited by 4,100
Research works printed but not linked: 1
- Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the Association for Computational Linguistics, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics.
Cited by
Sorted by citations
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
On the Opportunities and Risks of Foundation Models
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021) · arXiv · doi:10.48550/arxiv.2108.07258 · Cited by 2,676
Language (Technology) is Power: A Critical Survey of “Bias” in NLP
Blodgett, S. L., Barocas, S., Daumé III, H., & Wallach, H. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.485 · Cited by 1,050
“…and Ungar (2019); Stanovsky et al. (2019); Tan and Celis (2019); Zhao et al. (2019); Zhou et al. (2019); Chopra et al. (2020); Gyamfi et al. (2020); Nadeem et al. (2020); Nissim et al. (2020); Papakyriakopoulos et al. (2020); Popović et al. (2020); Ravfogel et al. (2020); Ross et al. (2020);…”
Bias and Fairness in Large Language Models: A Survey
Gallegos, I. O., Rossi, R. A., Barrow, J., Tanjim, M. M., Kim, S., Dernoncourt, F., Yu, T., Zhang, R., & Ahmed, N. K. (2024) · Computational Linguistics · doi:10.1162/coli_a_00524 · Cited by 1,245
Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., & Fritz, M. (2023) · Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security · doi:10.1145/3605764.3623985 · Cited by 667
“LLMs might make up facts (“hallucinate”), generate polarized content, or reproduce biases, hate speech, or stereotypes [39, 41, 47, 57, 61, 65, 78].”
Background
Metadata
- Authors
- Nadeem, M., Bethke, A., & Reddy, S.
- Affiliations
- Massachusetts Institute of Technology (US)
- University of California, Santa Barbara (US)
- University of Cambridge (GB)
- ORCID
- Reddy, Siva
- Venue
- Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)
- Published
- 2021
- Pages
- 5356-5371
- OpenAlex
- W3019416653
- Semantic Scholar
- babeda48b10a4d638252118f2238d05a06f4ec55
- Cited by
- 643
- References
- 15
- Citations by year
- 2026: 70 · 2025: 131 · 2024: 98 · 2023: 191 · 2022: 85 · 2021: 49
- FWCI
- 95.35
- Type
- conference paper
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- Locations
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Hate Speech and Cyberbullying Detection
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- stereotypical bias, pre-trained language models, gender stereotypes, language model evaluation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-23, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-23 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
