TIDEs: A Transgender and Nonbinary Community-Labeled Dataset and Model for Transphobia Identification in Digital Environments

It may state two separate claims as one.

Using the toxicity metric on the best-performing threshold (0.1), the Perspective API attained an F1 score of 0.72, with precision of 0.91, and recall of 0.59.

“Using the toxicity metric on the best-performing threshold (0.1), the Perspective API attained an F1 score of 0.72, with precision of 0.91, and recall of 0.59.”

It may state two separate claims as one.

Casual transphobic rhetoric was most concentrated in the YouTube data, though it was found across all data sources.

“Casual transphobic rhetoric was most concentrated in the YouTube data, though it was found across all data sources.”

It may state two separate claims as one.

The dataset includes sub-codes: anti-queer (n = 16), bio-essentialist (n = 15), uninformed (n = 15), radical feminist (n = 11), and sex-based language (n = 11).

“We descriptively coded these instances during consensus meetings with sub-codes, arriving at the following categories: anti-queer (n = 16), bio-essentialist (n = 15), uninformed (n = 15), radical feminist (n = 11), and sex-based language (n = 11).”

It may state two separate claims as one.

The majority of these racist posts were anti-Black and a smaller number were anti-immigrant.

“The majority of these posts were anti-Black and a smaller number were anti-immigrant.”

It may state two separate claims as one.

The model performs best on Tumblr and Truth Social data and worst on Reddit and YouTube data, especially on recall.

“The model performs best on Tumblr and Truth Social data and worst on Reddit and YouTube data, especially on recall.”

It may state two separate claims as one.

Using the identity attack metric on the best-performing threshold (0.1), the Perspective API attained an F1 score of 0.74, with precision of 0.80, and recall of 0.68.

“Using the identity attack metric on the best-performing threshold (0.1), the Perspective API attained an F1 score of 0.74, with precision of 0.80, and recall of 0.68.”

It may state two separate claims as one.

The Logistic Regression model attained an F1 score of 0.77, with precision of 0.87, and recall of 0.70.

“The Logistic Regression model attained an F1 score of 0.77, with precision of 0.87, and recall of 0.70.”

It may state two separate claims as one.

The final TIDEs dataset contains 3,479 posts, with 54.8% not transphobic (n = 1,907) and 45.1% transphobic (n = 1,572).

“The final TIDEs dataset (See Table 2, Figure 1) contains a balanced mix of posts across the two different classes, with 54.8% not transphobic posts (n = 1,907) and 45.1% transphobic posts (n = 1,572), for a total of 3,479 posts.”

It may state two separate claims as one.

Truth Social had the highest proportion of transphobic posts, while Reddit had the highest proportion of not transphobic instances.

“Truth Social had the highest proportion of transphobic posts, while Reddit had the highest proportion of not transphobic instances.”

It may state two separate claims as one.

TERF rhetoric was frequently found in the Tumblr data, and in some Reddit instances.

“Similarly, trans-exclusionary radical feminist (TERF) [8] rhetoric was frequently found in the Tumblr data, and in some Reddit instances.”

• cites [8]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

It may state two separate claims as one.

The fine-tuned DeBERTa model succeeds at detecting several ideologically distinct types of transphobia, achieving an F1 score of 0.81.

“Our fine-tuned DeBERTa model succeeds at detecting several ideologically distinct types of transphobia, achieving an F1 score of 0.81.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

The dataset contains many context-dependent posts that the authors identified as transphobic because of their relevance in current transphobic discourse.

“Our dataset contains many context-dependent posts that we identified as transphobic because of their relevance in current transphobic discourse.”

It may state two separate claims as one.

Many posts in the dataset, especially on Truth Social, were racist in addition to being transphobic, or were racist without being transphobic.

“While labeling the data, we found that many posts, especially on Truth Social, were racist in addition to being transphobic, or were racist without being transphobic.”

It may state two separate claims as one.

The dataset contains difficult-to-label posts that are overtly anti-men or anti-queer.

“There were other types of difficult-to-label posts, such as posts that are overtly anti-men or anti-queer (as a word and identity).”

It may state two separate claims as one.

The DeBERTa model achieved an F1 score of 0.81, with precision of 0.89 and recall of 0.75.

“The DeBERTa model achieved an F1 score of 0.81, with precision of 0.89 and recall of 0.75.”

It may state two separate claims as one.

The Perspective API was able to identify some transphobia when set to a very low threshold, though it performed worse than the TIDEs model.

“As a comparative baseline, the Perspective API was able to identify some transphobia when set to a very low threshold, though it performed worse than the TIDEs model.”

It may state two separate claims as one.

The model detects many kinds of transphobia but still struggles to detect transphobia in some subtle or harder to parse instances.

“Since our dataset included a very wide range of rhetoric, the model detects many kinds of transphobia, but it still struggles to detect transphobia in some subtle (row 2) or harder to parse (row 1) instances.”

It may state two separate claims as one.

The dataset includes data from right-wing transphobes, TERF transphobes, and casual transphobes.

“Through this process, we were able to include data from right-wing transphobes, TERF transphobes, and casual transphobes.”

It may state two separate claims as one.

There were some false positives on shorter posts with less context and on harder to parse posts.

“Conversely, there were some false positives on shorter posts with less context (row 4) and similarly on harder to parse posts.”

It may state two separate claims as one. The quote only partly states this claim.

TIDEs incorporates data from Tumblr, Truth Social, YouTube, and Reddit and can identify a combination of different kinds of transphobia from a range of different social media sources.

“To further ongoing efforts to mitigate transphobia online, we created the TIDEs (Transphobia Identification in Digital Environments) dataset, which incorporates data from Tumblr, Truth Social, YouTube, and Reddit, and used it to fine-tune models for transphobic detection.”

It may state two separate claims as one.

Logistic Regression achieved accuracy 0.79, precision 0.87, recall 0.70, macro F1 0.79, and weighted F1 0.77 on recognizing transphobic posts.

“| Logistic | 0.79 0.87 0.70 | 0.79 | 0.77 |”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

The YouTube subset contains 135 transphobic and 592 not transphobic posts (total 727).

“| Youtube | 135 | 592 | 727 |”

It may state two separate claims as one.

Perspective API (toxicity) achieved accuracy 0.63, precision 0.91, recall 0.59, macro F1 0.60, and weighted F1 0.72 on recognizing transphobic posts.

“| Perspective API | 0.63 0.91 0.59 | 0.60 | 0.72 |”

It may state two separate claims as one.

In cross-source testing, the best model trained on the other three sources achieved accuracy 0.27, macro F1 0.21, weighted F1 0.12, and F1 0.43 on Reddit.

“| Reddit | 0.27 | 0.21 | 0.12 | 0.43 |”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

In the logistic regression model, the token 'transgender' had the highest positive coefficient (2.91) for the transphobic class.

“| transgender | 2.91 | good | -1.08 |”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

The Reddit subset contains 123 transphobic and 766 not transphobic posts (total 889).

“| Reddit | 123 | 766 | 889 |”

Probably not a substantial claim. The authors may not put it forward as true. The paper’s conclusions may not depend on it.

In the logistic regression model, the token 'good' had the strongest negative coefficient (-1.08) for the not transphobic class.

“| transgender | 2.91 | good | -1.08 |”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

The Tumblr subset contains 464 transphobic and 409 not transphobic posts (total 873).

“| Tumblr | 464 | 409 | 873 |”

It may state two separate claims as one.

In cross-source testing, the best model trained on the other three sources achieved accuracy 0.53, macro F1 0.53, weighted F1 0.52, and F1 0.48 on Tumblr.

“| Tumblr | 0.53 | 0.53 | 0.52 | 0.48 |”

It may state two separate claims as one.

The dataset contains 3479 total posts, with 1572 labeled transphobic and 1907 labeled not transphobic.

“| Total | 1572 | 1907 | 3479 |”

It may state two separate claims as one.

In cross-source testing, the best model trained on the other three sources achieved accuracy 0.71, macro F1 0.63, weighted F1 0.69, and F1 0.46 on YouTube.

“| YouTube | 0.71 | 0.63 | 0.69 | 0.46 |”

It may state two separate claims as one.

DeBERTa achieved accuracy 0.82, precision 0.89, recall 0.74, macro F1 0.82, and weighted F1 0.81 on recognizing transphobic posts.

“| DeBERTa | 0.82 0.89 0.74 | 0.82 | 0.81 |”

It may state two separate claims as one.

Perspective API (identity attack) achieved accuracy 0.71, precision 0.80, recall 0.68, macro F1 0.70, and weighted F1 0.74 on recognizing transphobic posts.

“| Perspective API | 0.71 0.80 0.68 | 0.70 | 0.74 |”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

Even if TIDEs is used in the future to create a novel tool that helps protect trans social media users from toxicity by hiding it or flagging it through an interface, the transphobic content in question would still be present on the platform.

“As such, even if TIDEs is used in the future to create a novel tool that helps protect trans social media users from toxicity by hiding it or flagging it through an interface, the transphobic content in question would still be present on the platform.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper may give no reason why it holds. It may state two separate claims as one.

New transphobia campaigns will continue to emerge, and with them new language that the current dataset is not trained on.

“New transphobia campaigns will continue to emerge, and with them new language that our current dataset is not trained on.”

It may state two separate claims as one.

By reading each of these instances together and discussing them at length, the authors reached a carefully constructed final dataset that annotators fully agreed on, with additional comments on particularly difficult instances.

“By reading each of these instances together and discussing them at length, we reached a carefully constructed final dataset that annotators fully agreed on, with additional comments on particularly difficult instances.”

Probably not a substantial claim. It may state two separate claims as one.

As a result of the care-centered approach, the authors understand transphobic rhetoric online as a multifaceted issue that can manifest in different forms and contexts, and they capture these in the TIDEs dataset.

“As a result of our care-centered approach, we understand transphobic rhetoric online as a multifaceted issue that can manifest in different forms and contexts, and we capture these in the TIDEs dataset.”

Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.

Anti-men and anti-queer sentiments are often expressed by transphobic people, in both directly and indirectly transphobic ways.

“Neither of these sentiments is necessarily transphobic, but anti-men and anti-queer sentiments are often expressed by transphobic people, in both directly and indirectly transphobic ways.”

It may state two separate claims as one.

The authors' approach extends prior work that separated these tasks by joining them into one, thus making it possible to detect complex and varied transphobic speech.

“Our approach extends prior work that has separated these tasks [15,47] by joining them into one, thus making it possible to detect complex and varied transphobic speech.”

• cites [15]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [47]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

It may state two separate claims as one.

TIDEs can detect a high percentage of transphobic rhetoric, indicating its potential for being deployed in limited settings where content is flagged rather than censored.

“TIDEs can detect a high percentage of transphobic rhetoric, which indicates its potential for being deployed in limited settings where content is flagged rather than censored.”

Probably not a substantial claim. It may be too vague to check against evidence.

Moving toward community-centered labeling and moderation practices is crucial for creating systems that can last beyond a few months and be used in real-world contexts.

“Moving toward community-centered labeling and moderation practices is crucial for creating systems that can last beyond a few months and be used in real-world contexts.”

Probably not a substantial claim. It may be too vague to check against evidence.

The authors' work is a step toward developing trans technologies that can work for trans people and address transphobic content online.

“Ultimately, our work is a step toward developing trans technologies that can work for trans people and address transphobic content online.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

The dataset was split into train, development, and test sets with an 80:10:10 ratio.

“After finalizing the dataset, we split it into train, development, and test sets with an 80:10:10 ratio.”

It may state two separate claims as one.

The authors' model performance metrics demonstrate that a single classifier is able to detect transphobic rhetoric from a variety of different sources and actors, highlighting the potential for it to be deployed to address transphobia across social media contexts, with some limitations.

“Our model performance metrics demonstrate that a single classifier is able to detect transphobic rhetoric from a variety of different sources and actors, highlighting the potential for it to be deployed to address transphobia across social media contexts, with some limitations.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

The authors, as anti-racist researchers, struggled to conceive of a transphobia classifier that they would use in their daily lives that would flag transphobia and not white supremacy.

“As anti-racist researchers, we struggled to conceive of a transphobia classifier that we would use in our daily lives that would flag transphobia and not white supremacy.”

Probably not a substantial claim.

Care-centered data practices allowed the authors to revisit commonly-used methods in data work and consider how they could be adapted to better support trans people in the context of transphobia detection.

“Care-centered data practices allowed us to revisit commonly-used methods in data work and consider how they could be adapted to better support trans people in the context of transphobia detection.”

Probably not a substantial claim.

The authors introduce the concept of care-centered data practices, which span hiring and working with annotators, collecting data, and labeling data.

“Drawing from our process creating TIDEs we introduce the concept of care-centered data practices, which span hiring and working with annotators, collecting data, and labeling data.”

It may state two separate claims as one.

The DeBERTa model outperformed the Logistic Regression model, indicating that in some contexts transphobia is covert enough that purely lexical cues in a bag-of-words model do not sufficiently describe it.

“The DeBERTa model outperformed the Logistic Regression model, indicating that, in come contexts, transphobia is covert enough that the purely lexical cues in a bag-of-words model do not sufficiently describe it.”

Probably not a substantial claim.

The authors annotated difficult instances with sub-codes to better capture the data's complexity.

“We annotated difficult instances with sub-codes to better capture the data's complexity.”

Probably not a substantial claim. The paper may give no reason why it holds.

Trans people often create or adopt technologies independent from (but deployed within) platforms that help them mitigate the effects of facing transphobia online.

“As such, trans people often create or adopt technologies independent from (but deployed within) platforms that help them mitigate the effects of facing transphobia online.”

Probably not a substantial claim. It may be too vague to check against evidence.

Without acknowledging the complexity and variety of transphobic rhetoric online, the authors would risk creating a classifier that may perform well in tests but cannot be applied in social media contexts to detect non-extreme forms of transphobia.

“Without acknowledging the complexity and variety of transphobic rhetoric online, we would risk creating a classifier that may perform well in tests, but that cannot be applied in social media contexts to detect non-extreme forms of transphobia.”

Probably not a substantial claim.

The care-centered data practices described in the paper were crucial to how the authors approached the task of creating a transphobia classifier.

“The care-centered data practices described in this paper were crucial to how we approached the task of creating a transphobia classifier.”

Probably not a substantial claim. The paper may give no reason why it holds.

The care-centered practices contributed to a rich dataset that surfaces interesting and useful data without sacrificing annotator well-being.

“This contributed to a rich dataset that surfaces interesting and useful data without sacrificing annotator well-being.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

Checking in frequently and working closely with annotators was crucial for safeguarding their well-being.

“Checking in frequently and working closely with annotators was crucial for safeguarding their well-being.”

It may state two separate claims as one.

Annotators were meaningfully involved throughout the entirety of the project and understood the goal of the dataset and classifier and how their annotations would be used.

“Annotators were meaningfully involved throughout the entirety of the project; they understood the goal of the dataset and classifier and how their annotations would be used.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.

Transphobic discourse online is a manifestation of transphobia as a structural form of oppression.

“Additionally, transphobic discourse online is a manifestation of transphobia as a structural form of oppression.”

• AssumptionRests on an assumption: Transphobic discourse online is a manifestation of transphobia as a structural form of oppression.

Probably not a substantial claim.

The authors contribute a perspective on data annotation that makes use of qualitative coding methods and prioritizes careful consideration of trans people's varying opinions over arriving at a falsely cohesive understanding of what is and is not transphobic.

“As such, we contribute a perspective on data annotation that makes use of qualitative coding methods (e.g., [63]) and prioritizes careful consideration of trans people's varying opinions over arriving at a falsely cohesive understanding of what is and is not transphobic.”

• cites [63]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim. It may be too vague to check against evidence.

Further iteration on the dataset and model would be needed before it could be deployed for content moderation tasks that remove or hide content.

“Further iteration on the dataset and model would be needed before it could be deployed for content moderation tasks that remove or hide content.”

Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

The codebook was shared with annotators, who were encouraged to contribute to it as they annotated.

“After hiring data annotators, we shared this version of the codebook with them, clarifying that it was not a comprehensive list of all types of transphobia, and encouraged them to contribute to the codebook as they annotated.”

It may state two separate claims as one.

Data annotators are often not considered part of a research team and are excluded from peer support networks.

“While prior work has emphasized the importance of building peer support networks among researchers [16], we highlight that data annotators are often not considered part of a research team, and as such they are excluded from these networks.”

• cites [16]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.

Transphobic rhetoric is a prevalent problem on social media that existing platform policies fail to meaningfully address.

“Transphobic rhetoric is a prevalent problem on social media that existing platform policies fail to meaningfully address.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.

Deploying technological tools for transphobia detection is a limited response to a problem that cannot be simply addressed through technology.

“While we believe that there is value to deploying technological tools for transphobia detection, we acknowledge that this is a limited response to a problem that cannot be simply addressed through technology.”

Probably not a substantial claim.

The authors' priority was to ensure that annotators would be safe and comfortable.

“While we believe that our care-centered data practices had a positive impact on the data, resulting in a carefully constructed dataset, our priority was to ensure that annotators would be safe and comfortable.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.

The care-centered data practices had a positive impact on the data, resulting in a carefully constructed dataset.

“While we believe that our care-centered data practices had a positive impact on the data, resulting in a carefully constructed dataset, our priority was to ensure that annotators would be safe and comfortable.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.

The authors' practices in developing and releasing the TIDEs dataset and model reflect their commitment to conducting research that centers and benefits trans social media users.

“Our practices in developing and releasing the TIDEs dataset and model reflect our commitment to conducting research that centers and benefits trans social media users.”

Probably not a substantial claim. The paper may give no reason why it holds.

A transphobia classifier not trained on TERF data will fail to detect a significant amount of transphobic rhetoric that greatly impacts trans people.

“If a transphobia classifier is not trained on TERF data, for example, it will fail to detect a significant amount of transphobic rhetoric that greatly impacts trans people.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper may give no reason why it holds. It may state two separate claims as one.

Care for trans people entails reckoning with the different experiences and perspectives within the trans community, which in turn shape how different trans people understand transphobia.

“While many kinds of speech are undoubtedly transphobic, care for trans people entails reckoning with the different experiences and perspectives within the trans community, which in turn shape how different trans people understand transphobia.”

Probably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds. It may state two separate claims as one.

Spending additional cost and time to reach consensus on data annotation is not a standard practice in NLP hate speech classification tasks, but it allowed for deeper engagement with the data and the data annotators.

“Spending additional cost and time to reach consensus on data annotation is not a standard practice in NLP hate speech classification tasks, but it allowed for deeper engagement with the data and the data annotators.”

Probably not a substantial claim.

TIDEs makes a contribution to trans-centered content moderation efforts by serving as a novel tool for transphobia detection.

“TIDEs makes a contribution to trans-centered content moderation efforts by serving as a novel tool for transphobia detection which we hope to build on and share with other researchers.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

Prioritizing annotator well-being can help to mitigate the negative impacts of hate speech annotation.

“Prioritizing annotator well-being can help to mitigate the negative impacts of hate speech annotation [3].”

• cites [3]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.

One of the primary challenges with hate speech detection is the quick pace at which new coded language develops on social media.

“One of the primary challenges with hate speech detection is the quick pace at which new coded language develops on social media.”

Probably not a substantial claim. It may be too vague to check against evidence.

A useful transphobia detection tool must be developed with an awareness of the idiosyncrasies of transphobia, so that non-conventional speech does not go undetected.

“Instead, we reckon with the fact that a useful transphobia detection tool must be developed with an awareness of the idiosyncrasies of transphobia, so that non-conventional speech does not go undetected.”

Probably not a substantial claim.

TIDEs as an intervention is focused on individual responses to transphobia in the face of a lack of protection from platforms.

“TIDEs as an intervention is focused on individual responses to transphobia in the face of a lack of protection from platforms.”

It may state two separate claims as one.

The TIDEs model succeeded in detecting transphobia in the dataset, outperforming the baselines.

“Overall, the TIDEs model succeeded in detecting transphobia in the dataset, outperforming our baselines.”

Probably not a substantial claim.

Throughout its development, the authors adopted practices that allowed them to engage with the data and annotators through a lens of care for trans people.

“Throughout its development, we adopted practices of that allowed us to engage with the data and annotators through a lens of care for trans people.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.

TIDEs qualifies as a kind of trans technology.

“By creating TIDEs, we advance work that creates safer online spaces for trans people; thus, TIDEs qualifies as a kind of trans technology.”

Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

The authors strive to achieve balance between addressing the root causes of prevalent issues such as transphobia while also developing technology that works to alleviate or address the immediate effects of transphobic speech online.

“We strive to achieve balance between addressing the root causes of prevalent issues such as transphobia while also developing technology that works to alleviate or address the immediate effects of transphobic speech online.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

Annotators were explicitly told that they could choose to stop working at any point with no penalty.

“We explicitly told annotators that they could choose to stop working at any point with no penalty.”

Probably not a substantial claim.

The prevalence of racist rhetoric in the dataset foregrounds the importance of understanding transphobia in conjunction with racism.

“These posts, as well as the general prevalence of racist rhetoric in the dataset, foreground the importance of understanding transphobia in conjunction with racism.”

Probably not a substantial claim.

Research teams may not consider the impact that hate speech annotation may have on annotators.

“Research teams may not consider the impact that hate speech annotation may have on annotators, not only because annotators are not considered part of the research team, but also because there is typically very little interaction between researchers and annotators who are hired on platforms like Amazon Mechanical Turk.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

The community labeling feature is a form of trans-centered moderation.

“This is a form of trans-centered moderation [69], and a research space that we hope will continue to grow.”

• cites [69]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim.

The careful process of creating TIDEs foregrounded the context- and reader-dependent nature of transphobic rhetoric.

“The careful process of creating TIDEs foregrounded the contextand reader-dependent nature of transphobic rhetoric.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

Annotators were encouraged to space out labeling across several sessions and take frequent breaks.

“Annotators were encouraged to space out labeling across several sessions and take frequent breaks [52].”

• cites [52]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim.

The authors resolved discrepancies in labeling between annotators by holding consensus meetings where the two annotators discussed each specific case where their labels differed.

“We resolved discrepancies in labeling between annotators by holding consensus meetings, where the two annotators who labeled each portion of the dataset discussed each specific case where their labels differed.”

Probably not a substantial claim.

The paper outlines care-centered data practices, a methodology for constructing and labeling datasets for hate speech classification, developed while working closely with trans and nonbinary data annotators.

“We outline care-centered data practices, a methodology for constructing and labeling datasets for hate speech classification, which we developed while working closely with trans and nonbinary data annotators.”

Probably not a substantial claim. It may be too vague to check against evidence. It may state two separate claims as one.

The TIDEs system can be improved upon, iterated on, and used to develop tools to address transphobia detection without relying on companies or services that may become hostile against trans people in the future.

“This system can be improved upon, iterated on, and used to develop tools to address transphobia detection without relying on companies or services that may become hostile against trans people in the future.”

Probably not a substantial claim. It may be too vague to check against evidence.

The three types of care-centered data practices positively shaped how the authors engaged with the process of transphobia detection and allowed for a more thoughtful project for the benefit of trans communities.

“In conjunction, these three types of care-centered data practices (while hiring and working with data annotators, while collecting data, and while labeling data) positively shaped how we engaged with the process of transphobia detection, and allowed for a more thoughtful project for the benefit of trans communities.”

Probably not a substantial claim. It may state two separate claims as one.

The authors' work is grounded both in a situated understanding of what makes transphobia complex and difficult to detect and a perspectivist approach to data annotation and hate speech classification.

“Our work is grounded both in a situated understanding of what makes transphobia complex and difficult to detect and a perspectivist approach to data annotation and hate speech classification [25].”

• cites [25]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim. The paper’s conclusions may not depend on it.

Before the data annotation process began, the authors created a preliminary codebook that identified basic examples of common transphobic rhetoric.

“Before the data annotation process began, we created a preliminary codebook that identified basic examples of common transphobic rhetoric (e.g., misgendering, calling trans people mentally ill).”

Probably not a substantial claim. It may be too vague to check against evidence.

This trans-centered research is, first and foremost, about creating tangible tools that can help trans people experience a less toxic internet.

“This trans-centered research is, first and foremost, about creating tangible tools that can help trans people experience a less toxic internet.”

Probably not a substantial claim. It may be too vague to check against evidence.

TIDEs can serve as the base for future trans technologies and research that confronts and addresses online transphobia.

“As a publicly available dataset and model, TIDEs can serve as the base for future trans technologies and research that confronts and addresses the problem of online transphobia.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.

It is crucial that researchers be aware of the negative impacts that conducting difficult research may have on them.

“As Andalibi and Forte [4] have discussed, it is crucial that researchers be aware of the negative impacts that conducting difficult research may have on them.”

• cites [4]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim.

The authors avoid thinking of transphobia as a form of bigotry that is simply expressed and easily recognizable.

“We avoid thinking of transphobia as a form of bigotry that is simply expressed and easily recognizable.”

Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.

The authors' data collection decisions are practices of care, because they approach this research problem in a way that accounts for the many complex ways transphobia emerges and affects trans people.

“These data collection decisions are practices of care, because we approach this research problem in a way that accounts for the many complex ways transphobia emerges and affects trans people.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

Prior work has emphasized the importance of building peer support networks among researchers.

“While prior work has emphasized the importance of building peer support networks among researchers [16], we highlight that data annotators are often not considered part of a research team, and as such they are excluded from these networks.”

• cites [16]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.