“She’ll Never Be a Man” A Corpus-Based Forensic Linguistic Analysis of Misgendering Discrimination on X

Abstract

Misgendering is a form of microaggression that reinforces gender binarism and involves the use of incorrect pronouns, names or gendered language when referring to a transgender and gender non-conforming (TGNC) individual. Despite growing awareness, it remains a persistent form of discrimination, and it is crucial not only to understand and address misgendering but also to analyse its impact within online discourse towards the TGNC community. The present study examines misgendering directed at the TGNC community present on platform X. To achieve this, a representative sample of 400 tweets targeting two TGNC individuals is compiled, applying an annotation scheme to manually classify the polarity of each tweet and instances of misgendering, and then comparing the manual annotations with those of an automatic sentiment detection system. The analysis focuses on the context and frequency of intentional misgendering, using word lists to examine the data. The results confirm that misgendering perpetuates discrimination, tends to co-occur with other forms of aggression, and is not effectively identified by automatic sentiment detection systems. Finally, the study highlights the need for improved automatic detection systems to better identify and address misgendering in online discourse and provides potentially useful tools for future research.

1. Introduction

Gender identity is a fundamental aspect of the human being that reflects the inner sense of who we are. Recognition and respect for gender identity are essential to build an inclusive society, where each individual can live following their true identity without fear of reprisal or discrimination.

Transgender and gender non-conforming (hereafter TGNC) individuals represent a complex and diverse population whose gender identities defy conventional societal norms (Argyriou 2021, p. 71). Rooted in a deep sense of self-awareness and authenticity, these individuals experience a profound misalignment between their inner sense of gender and the sex they were assigned at birth (American Psychological Association 2015).

Furthermore, within the TGNC community, a rich tapestry of experiences and identities exists. Some individuals choose to undergo physical transitions, such as hormone therapy or surgical procedures, to align their bodies with their gender identity. In contrast, others opt for social transitions, changing their name, pronouns and outward presentation to reflect their true selves (Argyriou 2021, p. 71).

Despite this diversity within the TGNC community, individuals often face challenges in how they are perceived and, consequently, addressed by others. An example of this is the linguistic phenomenon of ‘misgendering’, a concept that has attracted increasing attention from scholars and the media, and that refers to the act of using incorrect pronouns or gendered language when referring to TGNC individuals (McLemore 2016). According to Argyriou (2021), “everyday life of TGNC people is filled with examples of invalidations of the kind, as misgendering is generalised and persistent” (p. 72). This can manifest in overt forms of disrespect, such as deliberate misidentification or verbal harassment, as well as more subtle forms of invalidation, such as dismissing a person’s gender identity or ignoring their preferred pronouns (Argyriou 2021, p. 73), who can sometimes appear together.

The consequences of these forms of disrespect, either overt or subtle, when performed repeatedly during a long period, are profound and have severe implications of cultivating a hostile atmosphere that can adversely affect the mental and emotional well-being of the targeted individuals. Moreover, when these acts become the norm, they reinforce harmful stereotypes and perpetuate systemic discrimination against TGNC people. They can have, as a final consequence, the neglect of access to mainstream society, which is the ultimate goal of hate speech (Guillén-Nieto 2023).

Hence, it is necessary to gain a deep understanding of the complexities surrounding misgendering, as it involves delving into the context of several different TGNC populations’ experiences. Each TGNC person’s journey is unique and shaped by various factors, including culture, background, and personal beliefs. Therefore, the attempt to understand and identify instances of misgendering faces a series of difficulties, which leads to the creation of the present study.

The primary aim of this study is to disseminate the concept of intentional misgendering, as a manifestation of discrimination expressed through language, to facilitate its subsequent detection on social networks, specifically platform X1. To accomplish this goal, a series of more specific objectives will be pursued:

• Create a dataset of tweets targeting TGCN individuals from platform X, applying a set of criteria to ensure relevance and accuracy.
• Implement an annotation scheme to classify each tweet’s polarity and evaluate the consistency of the manual annotation between two annotators.
• Examine the context and frequency of those tweets that include intentional misgendering, analysing wordlists of the instances.
• Evaluate the effectiveness of an automatic sentiment detection system by comparing its performance to manual annotations.
• Provide recommendations for improving automatic detection systems and addressing intentional misgendering in online discourse effectively.

By pursuing these interconnected objectives, this study aims to expand the understanding of the linguistic phenomenon of intentional misgendering, ultimately contributing to the creation of safer and more inclusive online environments for all users.

2. Theoretical Background

Language, as a fundamental communication tool, can be weaponised to inflict harm, with words carrying both damaging and legally significant consequences. Forensic linguistics is key in examining how such harmful language can serve as evidence in legal contexts (Guillén-Nieto 2022, p. 1). In this context, this section offers an in-depth review of essential concepts central to understanding harmful language, starting with the broader framework of harassment, progressing through microaggressions, and ultimately focusing on misgendering, which is the core of this study. The purpose is to clarify how these forms of derogatory language interrelate and provide a more comprehensive understanding of their objectives and their legal and social implications.

2.1. The Concept of Harassment

Harassment is a highly complex and multifaceted phenomenon that encompasses a wide range of offensive behaviours designed to undermine an individual’s dignity. These behaviours, which often consist of hostile and unethical forms of communication, are directed by one or multiple perpetrators victimising a particular target for an extended period (Leymann 1990). According to Guillén-Nieto (2022, p. 7), it involves a series of acts directed towards the destruction or diminution of the fundamental rights of the affected individual. The objective behind such actions is typically malicious, mainly due to the desire of perpetrators to achieve certain aims or goals, which may vary depending on the context. For instance, the case of gender-based harassment faced by the TGNC community contributes to the central goal of hate speech, which is to deny them equal access and exclude them from the rest of society. This type of harassment, considered a macro-directive, is carried out through the execution of a series of micro-acts of aggression, each of which indirectly contributes to the achievement of these super-goals (Guillén-Nieto 2022, p. 8). Hence, the series of smaller aggressive actions conforming to harassment receive the name of ‘microaggressions’.

2.2. The Concept of Microaggressions

In 1969, Dr Chester Pierce introduced the term ‘offensive mechanisms’ to describe the subtle but pervasive ways in which black people were marginalised in the United States. Pierce noted, “To be black in the United States today means to be socially minimised. Every day, black people face ‘offensive mechanisms’ designed to isolate, diminish, and confine them to a lesser status. The relentless message they receive is that they are unimportant and irrelevant” (Pierce 1970, p. 303). In a 1970 essay titled “Offensive Mechanisms”, Dr Pierce further developed this concept, coining the term ’microaggressions’ to refer to these understated yet impactful actions.

Additionally, psychologist Sue (2010) defined microaggressions as “brief, everyday interactions that convey negative or derogatory messages to people because of their identity group” (p. 36). Over time, the term has come to include not just racial bias but also insults and behaviours targeting other marginalised groups, such as ethnic minorities, gender minorities and people with disabilities (Sue 2010; Paludi 2012).

When targeting the TGNC community, the subtlety of microaggressions does not diminish their impact since it can lead to significant repercussions, including undermining a person’s identity and invalidating their existence. Therefore, the TGNC individuals are faced with a profound dissonance between their gender identity and the sex they were assigned at birth, which makes them vulnerable to misgendering—a form of microaggression that involves being addressed or named in a manner that is inconsistent with their gender identity (Argyriou 2021, p. 72).

2.3. The Concept of Misgendering

The American Psychological Association, in its “Guidelines for Psychological Practice with Transgender and Gender Nonconforming People,” defines ‘misgendering’ as “Using pronouns or other words that label a person’s gender incorrectly. This is often a painful experience for people including trans and gender nonconforming people, especially when done by someone aware of their gender identity.” (American Psychological Association 2015). This phenomenon serves as a potent reminder of the broader societal issues surrounding gender identity and acceptance and is crucial to combat this harmful practice with education, research and awareness. Thus, this study highlights misgendering as a central issue.

3. State of the Art

The literature review presents a detailed examination of research on ‘microaggressions’ and ‘misgendering’ from a linguistic perspective, to compile and present prior studies on the analysis and annotation of gender microaggressions. This review is divided into two main sections: descriptive linguistics and corpus linguistics approaches to gender microaggressions, focusing particularly on intentional misgendering.

3.1. Descriptive Linguistics Approach to Gender Microaggressions

Microaggressions have traditionally been examined in the context of racial and ethnic discrimination (Chang and Chung 2015, p. 220). Sue and Capodilupo (2008) were the first to identify parallels between racial and gender microaggressions, suggesting that the mechanisms underlying these biases might share commonalities.

In addition, Solórzano et al. (2000) was the first to coin the term ‘gender microaggressions’, but there has been limited empirical research to substantiate the concept. However, only in recent years has there been a significant expansion in the scope of research, encompassing microaggressions directed toward the LGBTQIA+ community. This broader focus reflects an increasing awareness of how microaggressions manifest, highlighting the need for further empirical investigations to understand and address these subtle but impactful forms of discrimination.

According to Nadal et al. (2016), research on microaggressions within the lesbian, gay and bisexual (LGB) communities has increased, yet studies addressing TGNC individuals remain sparse. This shortfall is partly due to the common conflation of gender identity with sexual orientation in broader LGBTQIA+ studies, which obscures the unique experiences of transgender people who face different forms of discrimination and marginalisation compared to their LGB counterparts (Fassinger and Arseneau 2007; McCarthy 2003). Despite inclusive intentions, combining these identities can inadvertently perpetuate the marginalisation of TGNC voices.

Sue et al. (2008) developed a classification system for gender-based microaggressions, which was later updated by Nadal et al. (2010). Building upon this work, Nadal et al. (2012) conducted a qualitative study that explored the nuanced interpersonal and systemic microaggressions faced by TGNC individuals. Participants from diverse backgrounds were recruited through local LGBTQIA+ organisations to form two focus groups. The study aimed to validate the existing taxonomy of microaggressions towards transgender people through directed content analysis, revealing twelve themes specific to TGNC individuals, thus expanding the understanding of their experiences.

Some of the themes identified included “the use of transphobic or incorrectly gendered terminology”, which involves derogatory terms or incorrect pronouns; “the assumption of a universal transgender experience”, which stereotypes transgender individuals; or “the exoticisation”, where transgender individuals are fetishised.

Nadal et al. (2016) observed that one of the most prevalent forms of microaggression encountered by TGNC individuals is the failure to recognise or consistently use their preferred pronoun, “particularly after someone has been corrected or informed of a genderqueer person’s preferences” (p. 13). Given this, the present study will concentrate on the first category identified in the taxonomy—namely, the use of transphobic or incorrectly gendered terminology—focusing on the phenomenon of ‘misgendering’. Misgendering involves the erroneous attribution of gender to an individual. This misattribution can materialise through the use of pronouns, titles or descriptions (McNamarah 2021). Moreover, the repercussions of such a practice can be highly damaging and deeply detrimental for those who experience it, as it undermines their sense of identity and belonging.

Particularly, this study focuses on ‘intentional misgendering’, which is characterised by a conscious and deliberate decision to disregard an individual’s preferred gendered language or titles. Unlike unintentional misgendering, which may stem from ignorance or oversight, intentional misgendering involves a deliberate choice to ignore or reject the correct gender designation of the person being addressed (McNamarah 2021, p. 2261).

In the literature on transgender identities and microaggressions, the phenomenon of misgendering has garnered significant attention. A recent study by Edmonds and Pino (2023) provided an in-depth analysis of intentional misgendering and its effects on power dynamics, identity construction and societal norms. Their research found that intentional misgendering is used to undermine transgender individuals’ identities, revealing how such acts are strategically employed to express negative attitudes and reinforce cisgenderism views. Trans individuals often respond by framing misgendering as morally reprehensible, challenging the discriminatory behaviour of the offenders.

Another study by Thál and Elmerot (2022) delved into the misgendering of transgender individuals in the Czech language, identifying linguistic infra-humanisation and dehumanisation as key components. The study highlighted the importance of recognising and addressing linguistic hostility towards transgender individuals through the analysis of diverse textual sources. This research underscores the need for sensitivity in language use to avoid reinforcing harmful practices and calls for greater awareness and inclusivity in communication to support TGNC individuals better.

However, despite the recent interest in this topic, there is still a gap in the literature, and this study seeks to fill it by exploring the phenomenon of intentional misgendering in English and collecting and annotating a corpus sample of tweets from platform X directed at TGNC individuals.

3.2. Corpus Linguistics Approach to Microaggressions Annotation

Within the framework of corpus linguistics, a significant gap is also identified in the literature regarding misgendering as a linguistic phenomenon. While misgendering, as a form of microaggression, has been subject to theoretical study for years, there is a notable absence of computational analyses on this topic. This gap represents a great opportunity to further explore and expand linguistic annotation efforts on the phenomenon of misgendering, which can serve the purpose of enriching language understanding and improving the inclusivity and accuracy of NLP technologies in various contexts.

Relevant work on gender bias annotation is the one by Havens et al. (2022), which addresses the challenge of mitigating gender bias in NLP systems by developing a taxonomy of gendered language and applying it to create annotated datasets. The taxonomy categorises labels into three main categories: Person Name, Linguistic, and Contextual, with sub-labels defined by archival documentation. The study underscores the significance of interdisciplinary collaboration and clear metrics in identifying gender bias and aims to guide NLP systems towards more inclusive representations of gender.

Assimakopoulos et al. (2020) conducted a study that introduces a hierarchical annotation scheme to discern discriminatory comments within online discussions. The scheme goes beyond the binary classification of hate speech versus non-hate speech, categorising comments based on their attitude towards target minority groups and detailing how negative attitudes are articulated. The study found that this multi-level scheme improved inter-annotator agreement, highlighting the need for refined annotation guidelines and comprehensive training for annotators to identify negative discourse strategies better.

These studies collectively highlight the importance of annotation in corpus-based research, particularly in identifying linguistic features and extracting insights from textual data. Annotation is crucial in advancing fields such as NLP by facilitating the systematic categorisation of elements within textual data. In the context of gender-based discrimination, effective annotation frameworks are essential for accurately capturing the complexity and impact of intentional misgendering.

The present research aims to fill the corpus collection and annotation gap related to misgendering. Developing robust annotation frameworks can enhance the detection of intentional misgendering, contributing to a more inclusive and respectful environment for TGNC individuals. This can help to address the root causes of misgendering and promote greater awareness of the importance of using correct and respectful language through focused corpus collection and annotation.

4. Research Questions

The present study formulates relevant research questions that will allow for a deeper investigation of the linguistic phenomenon presented in this study and the aspects that encompass it.

The questions proposed are as follows:

• RQ1: Does intentional misgendering as a form of microaggression perpetuate discrimination towards the TGNC community?
• RQ2: Does intentional misgendering typically co-occur with other forms of aggression or discriminatory language?
• RQ3: Is there a significant relationship between the presence of misgendering in tweets and their sentiment polarity?
• RQ4: Can automatic sentiment detection systems effectively identify tweets containing misgendering and expressing hatred towards transgender individuals, or is there a gap in their ability to detect this type of message?

These research questions will guide the compilation of messages extracted from platform X, as well as the evaluation of manual annotation and automatic polarity detection systems on misgendering directed at TGNC.

5. Methodology

5.1. Corpus Compilation

The corpus sample consists of a collection of tweets in English extracted from platform X, specifically focusing on intentional misgendering within the context of online discourse. This corpus sample aims to analyse and identify instances of intentional misgendering targeting two specific TGNC public figures. In addition, the text type selected for creating the corpus sample consists of tweets that mention two TGNC individuals who are public figures. This choice of text type is motivated by the prevalence of X as a platform for public discourse and its potential for capturing informal communication that reflects broader societal attitudes. Lastly, the selection of English as the language for the corpus sample is due to its widespread use on social media platforms and its relevance related to this Master.

5.1.1. Data Selection Criteria

To construct the corpus for studying intentional misgendering targeting the TGNC community, specific selection criteria have been established. In the words of Biber (1993), “a corpus must be ‘representative’ in order to be appropriately used as the basis for generalisations concerning a language as a whole” (p. 243). For this reason, the criteria selected to ensure the representativeness of the phenomenon in the corpus sample are the publication dates, size, authors and selection of the individuals targeted by misgendering.

Regarding the publication dates of the texts, tweets have been selected within a time range spanning from 1 January 2023, to the date of the compilation (15 April 2024). The choice of this period is based on significant changes in the “Abuse and Harassment Policies” of the social network X, particularly the removal and subsequent reintroduction of rules against misgendering between 2023 and 2024 (X 2024).

As for the sample size selection, it has been divided to represent both individuals equally. In total, 400 tweets were selected, with 200 tweets directed at each individual, creating a balanced framework for comparison and analysis, and ensuring representativeness. This sample size has been deliberately chosen to support the study’s qualitative approach, prioritising a deep, nuanced exploration of microaggressions to understand the complex nature of the phenomenon.

Considering the author of the corpus, it is not possible to establish specific criteria for the authors of the tweets since each tweet originates from different individuals with varied backgrounds, styles and motivations. Given this diversity, there is not a uniform set of criteria that can accurately categorise or classify them.

To further outline the corpus sample criteria, it is essential to explain the selection of the individuals targeted by misgendering. The choice is based on their public prominence and the controversies surrounding their transitions to ensure a certain number of tweets are directed at them.

In addition, it is important to remark that as Burghardt (2015) states, “redistributing Twitter content outside the Twitter platform is prohibited. In practice, this means it is not possible to precompile Tweet corpora and to share them in a way they are readily accessible for academic research.” (p. 78). This means that to carry out the corpus sample compilation, the individuals need to face an anonymisation process where their names are changed to hide their identity. To carry out the corpus sample compilation, the individuals must face an anonymisation process where their names are changed to hide their identity. However, misgendering is inherently contingent upon the identity, context and cultural setting within which it occurs, as noted by Hochdorn et al. (2016). For that reason, it is necessary to briefly contextualise the situation of the individuals to understand the reasoning behind the choice for this study:

• “Individual 1” is a transgender man famous for his work as an actor before his transition. His pronoun set is he/they.
• “Individual 2” is a transgender woman known for documenting her transition process on social media. Her pronoun set is she/they.

After establishing the criteria for the corpus sample, the next step in this stage involves extracting tweets that meet these criteria.

5.1.2. Data Extraction

The first approach to extracting tweet data involves using the Twitter Search API to access tweets from the ‘Top’ category. For this method, tweets containing the keywords “Individual 1” or “Individual 2” from 1 January 2023 to 15 April 2024 are retrieved. To enhance the data collection, a second approach using web scraping methods is also employed. This complementary method involves creating a query search on X and scraping the web code to extract each of the tweets one by one.

5.1.3. Data Pre-Processing

The raw data obtained from X underwent a pre-processing stage to refine it into a more usable format. Given that raw data often include numerous irrelevant tweets, a manual effort was made to simplify the dataset, thereby minimising the annotation of tweets not fitting the established criteria. To achieve this, the following filtering mechanisms devised to selectively extract tweets were employed:

• Keyword presence: Tweets explicitly mention the keyword “Individual 1” or “Individual 2”.
• Duplicate tweets: Excluding tweets that are duplicates.
• URL/image tweets: Filtering out tweets solely consisting of URLs/images.
• Language criterion: Eliminating tweets composed in languages other than English or those containing a mix of English and non-English content.
• Minimum length: Disregarding tweets with a character count below five.
• User mentions: Removing tweets primarily comprised of user2 mentions.

Thus, implementing this filtering process ensures the efficiency of subsequent annotation efforts and the dataset’s quality and integrity.

5.1.4. Dataset Statistics

The corpus sample is composed of 400 tweets, with an equal distribution of 200 tweets referring to Individual 1 and 200 referencing Individual 2. The complete corpus contains a total of 14,492 tokens and 12,284 unique types. However, to ensure a comprehensive examination of misgendering patterns, it was divided into two subsets or sub-corpus based on whether the tweets mention Individual 1 or 2. This division facilitates a targeted analysis of misgendering, recognising that its patterns may differ when directed at a trans man versus a trans woman. The first subset, comprising tweets referencing Individual 1, includes 7458 tokens and 6276 unique types, while the second subset, centred on Individual 2, contains 7034 tokens and 6008 unique types.

In addition, a wordlist frequency was generated for both subsets using Sketch Engine3. This platform provides a detailed breakdown of the common words and phrases from the corpus sample using the corpus enTenTen214 as a reference (Suchomel 2020). The following wordlists allow a more precise analysis of the language surrounding misgendering (see

Table 1. Wordlist of sub-corpus 1 from Sketch Engine.

No.	Lemma	Frequency	No.	Lemma	Frequency
1	Trans	26	11	Delusion	5
2	Lesbian	10	12	Tittie	5
3	Tit	9	13	Cisgender	3
4	Transitioned	8	14	Mutilate	3
5	Transphobe	7	15	Psychotic	3
6	Mastectomy	6	16	Deadnaming	2
7	Topless	5	17	Self-hatred	2
8	Cis	5	18	Weirdo	2
9	Slur	5	19	Transman	2
10	Trans	5	20	Objectify	2

and

Table 2. Wordlist of sub-corpus 2 from Sketch Engine.

No.	Lemma	Frequency	No.	Lemma	Frequency
1	Trans	18	11	Fiasco	3
2	Pretend	10	12	Womanhood	3
3	Gay	8	13	Manly	3
4	Dude	6	14	Backlash	3
5	Girlhood	4	15	Clown	3
6	Mock	4	16	Transgender	3
7	Pronoun	4	17	Influencer	3
8	Marginalise	4	18	Vomit-inducing	2
9	Leftist	4	19	Transvestite	2
10	Mockery	3	20	Transphobic	2

).

Overall, the combination of the corpus sample subsets and the wordlist frequency analysis provides a comprehensive foundation for understanding the complexities of misgendering in online discourse, laying the groundwork for further exploration and research in this critical area. These keyword tables will be used in the analysis of the results alongside the annotation data to examine the relationship between these terms and their associated sentiments, as well as to evaluate their connection to misgendering.

5.2. Corpus Annotation

After the data selection criteria, extraction and pre-processing procedures are completed, the second stage involves manually annotating the corpus sample. The upcoming sections detail the design of the annotation scheme and the steps taken in the annotation process to create a reliable and accurate corpus sample.

5.2.1. Annotation Scheme

The annotation scheme used in this corpus sample is designed to capture the sentiment expressed in the tweets while maintaining a consistent and reliable classification system. This scheme consists of three main polarity groups: neutral, negative or positive, which allows for a simple but comprehensive assessment of sentiment towards TGNC people in the extracted tweets. This methodology is outlined in the SemEval5 task by Rosenthal et al. (2015). Moreover, the simplicity and clarity of the selected annotation scheme contribute to reducing the ambiguity that could arise in such a context-dependent phenomenon as intentional misgendering. Some illustrative examples of tweets and their corresponding annotations are presented in

Table 3. Annotation of sentiment polarity in sub-corpus 1.

Polarity	Confidence	Tweets
Positive	1	I had an unusual dream about Individual 1 and now I think I might have feelings for him, lol.
Negative	1	She’ll never be a man, no matter what she does. Even though Individual 1 has invested a lot in trying to transition, people will not perceive her as part of our group.
Neutral	1	Film 1-directed by director 1, starring actor 1 and executive-produced by Individual 1-centers around a teenage girl navigating a competition.

and

Table 4. Annotation of sentiment polarity in sub-corpus 2.

Polarity	Confidence	Tweets
Positive	1	if Individual 2 did single-handedly disrupt the entire product 1 industry, she must be one of the most influential women in the world.
Negative	1	Individual 2: “I do not think God made an error with me.” He seems to be trying to sway people but inadvertently speaks the truth. Indeed, God didn’t make an error with him—he was made a man.
Neutral	1	Individual 2 was the winner of the first Woman of the Year award by brand 1.

. Per X’s anonymisation policies, the tweets have been paraphrased to ensure compliance.

5.2.2. Annotation Process

The annotation process involves several critical steps to ensure consistency and reliability. Initially, the tweets were assigned to two annotators with advanced Linguistics and English knowledge. Annotators were tasked with determining the overall polarity of each tweet based on their sentiment towards the targeted individuals, classifying them into one of three categories: neutral, negative or positive. Each annotator works independently to minimise bias and subjectivity during this classification stage.

Following the independent annotation, a reconciliation phase took place. During this phase, any discrepancies or disagreements in the sentiment classification are discussed and resolved through the final decision of a referee. This step is crucial to maintain consistency and ensure the annotations accurately represent the sentiment in the tweets.

Finally, after the reconciliation phase, the annotated corpus sample was compiled, along with a detailed record of the annotation decisions, including any resolved disagreements. This final corpus served as the foundation for further data analysis, with the annotations providing a structured framework for exploring sentiment and misgendering trends.

5.2.3. Annotation Guidelines

To ensure consistency and accuracy in the annotation process, specific guidelines have been established for annotators to ensure that the dataset is reliable and consistent. These guidelines are designed to help annotators focus on the annotation’s main objective, which is to determine the polarity of the tweets.

The first thing to consider when annotating sentiment, is whether the language used in the text respects or misrepresents the gender identity of the individuals being discussed. This involves examining the tone, word choice and contextual elements that indicate whether the sentiment is positive, neutral or negative. After determining the sentiment polarity, annotators should use a confidence scale from 0 to 1 to indicate the level of certainty for each annotation. This confidence measure helps to identify annotations that might require further review or discussion.

Furthermore, annotators are required to identify misgendering since it is the main research object in the present study. This procedure ensures all annotators can correctly recognise misgendering, providing a uniform understanding and approach to the annotation task.

5.2.4. Inter-Annotator Agreement

When annotating corpora involving multiple human annotators, it is critical to ensure consistency and reliability (Artstein and Poesio 2008, p. 556). This requirement is universal across all annotation types but is particularly crucial for corpus containing ambiguous or subjective factors such as intentional misgendering. This is because, with complex subjects, annotators might interpret annotation guidelines differently, leading to variation in their annotations. The inconsistency can defeat the purpose of the annotated corpus, as it can impede machine learning algorithms from extracting useful patterns for predictions. Thus, assessing the reliability of the annotation process is essential, using specific metrics to ensure high-quality outcomes (Moreno-Ortiz and García-Gámez 2022, p. 192).

To measure the consistency of the annotations, Cohen’s Kappa statistic (Cohen 1960, p. 42) was used. This metric accounts for the likelihood of agreement occurring by chance, providing a more robust measure of inter-annotator reliability than simple percentage agreement. In this instance, the kappa statistic was calculated on a subset of 100 tweets, out of 400, annotated by the two annotators. The resulting kappa value was 0.8168, indicating a high level of agreement between the annotators. This high kappa value suggests that the annotation process is consistent, providing confidence in the quality of the annotated data and supporting the reliability of subsequent analyses and machine learning applications.

6. Results and Discussion

This section presents the results obtained from the compilation and annotation of the corpus sample, providing a comprehensive analysis of the data to elucidate the nature and implications of intentional misgendering. The findings are quantified and exemplified to offer a detailed overview of the corpus.

The final annotation results provide valuable insights into two important aspects of the corpus: sentiment polarity and misgendering patterns. Sentiment polarity refers to whether tweets express a positive, negative or neutral sentiment towards the TGNC individuals. In contrast, intentional misgendering is annotated referring to the use or misuse of gender pronouns and gendered language to describe the two individuals. To start,

Table 5. Sentiment polarity of the corpus sample.

	Positive	Negative	Neutral	Total
Individual 1	57	126	17	200
Individual 2	26	153	21	200
Total	83	279	38	400

provides an overview of the tweets classified by the polarity of sentiment by the two annotators. The data suggest a strong negative bias in the sentiment of the overall corpus sample.

The aggregated data across both manual annotators reveal that negative tweets significantly outnumber positive and neutral ones, with 279 negative tweets compared to 83 positive and 38 neutral tweets. Notably, more than half of the tweets exhibit negative sentiments towards both individuals, underscoring the prevalence of derogatory content targeting TGNC individuals. These initial findings call for further research to explore the relationship between these sentiments and intentional misgendering.

Moreover, a more detailed analysis of misgendering was deemed necessary to determine whether this linguistic phenomenon differs based on the gender identity of the targeted individual. This approach aimed to identify and clarify any variations in how misgendering manifests when directed toward a trans man (Individual 1) versus a trans woman (Individual 2). Consequently, a sub-annotation process was conducted to categorise instances of intentional misgendering.

For this sub-annotation process, the data were analysed using a predefined set of classifications by McNamarah (2021) designed to capture the different forms of misgendering. By using these classifications, annotators could systematically identify and categorise instances of misgendering directed at both Individuals 1 and 2. This process aimed to provide a comprehensive understanding of misgendering in this context.

• The tweets that exhibit “mislabelling”, which involves using incorrect gendered terms or categories that do not align with the individual’s gender identity, are annotated as MISLABEL.
• The tweets that exhibit “mispronouning”, which entails using incorrect pronouns when addressing or referring to the individual, disregarding their gender identity, are annotated as MISPRONOUN.
• The tweets that use the correct pronouns or gendered language when referring to the individuals, aligning with their stated gender identity, are annotated as CORRECT GENDER.
• The tweets that do not directly address the individual’s gender identity and do not specify any gender-specific treatment are annotated as NO.

Other forms of misgendering, such as deadnaming, were excluded because the primary criterion for tweet inclusion was the use of the individuals’ chosen names post-transition. Consequently, ungendering and unpronouning were also excluded since both individuals have “they” as part of their pronoun set.

6.1. Analysis of the Results

The subsequent set of

Table 6. Annotation of misgendering in sub-corpus 1.

Misgendering	Positive	Negative	Neutral	Total
NO	18	20	9	47
CORRECT_GENDER	40	1	6	47
MISLABEL	0	27	1	28
MISPRONOUN	0	78	1	79
Total	57	126	17	200

and

Table 7. Annotation of misgendering in sub-corpus 2.

Misgendering	Positive	Negative	Neutral	Total
NO	17	49	18	84
CORRECT_GENDER	9	1	2	12
MISLABEL	0	33	1	34
MISPRONOUN	0	70	0	70
Total	26	153	21	200

aims to provide a more nuanced understanding of the polarity of sentiment in tweets, distinguishing between those containing instances of misgendering and those without. This division is achieved by detailing the specific type of misgendering used in each tweet.

For tweets where pronouns or gendered words targeting Individual 1 are absent (NO), the sentiment distribution was 20 negative tweets compared to 18 positive and 9 neutral. Among those that correctly used male pronouns and gender language (CORRECT GENDER), the distribution varied slightly, with 40 tweets that exhibited positive sentiment, only 1 tweet annotated as negative and 6 considered neutral.

For tweets containing mispronouning (MISPRONOUN), where pronouns like “she”, “her” or both together are used to refer to Individual 1, the data indicate a strong negative bias. Of the tweets with mispronouning, 78 are annotated as negative, with only 1 neutral. Regarding mislabelling (MISLABEL), the tweets were also predominantly negative, with 27 negative, 1 neutral and no positive tweets, and the words employed were mostly “woman” and “girl”.

Overall, the data suggest a correlation between misgendering and negative sentiment, with 126 tweets annotated as negatives, 78 presenting mispronouning and 27 with mislabelling. The high frequency of negative sentiment among tweets with misgendering indicates that such language is often used in a derogatory or hostile context when targeting Individual 1. Furthermore, the consistent pattern of misgendering with negative sentiment underscores the importance of further exploration to understand the underlying causes and the broader implications for TGNC individuals in social media discourse.

In addition, to assess the statistical validity of the data shown in Table 6, a Chi-Square test was conducted, resulting in a p-value < 0.05. This low p-value indicates that the differences observed in the data are not likely due to random chance, making the results statistically significant.

Additionally, when examining tweets targeting Individual 2, a similar pattern emerges concerning the relationship between misgendering and sentiment polarity. Tweets where pronouns or gendered terms are absent (NO) exhibit a negative sentiment bias, with 49 non-negative, 17 positive, and 18 neutral tweets.

In contrast, tweets that use correct female pronouns or gendered language (CORRECT GENDER) show a more positive distribution, with nine positive, one negative and two neutral tweets. However, compared to tweets directed at Individual 1, the number of tweets using the correct gender for Individual 2 is much lower. This could suggest a difference in how people refer to trans women compared to trans men, which may occur because of deep-rooted social perceptions and stereotypes that influence the way people use language to describe transgender individuals. Additionally, this discrepancy in using correct gender pronouns might also indicate that there is less public recognition or visibility for trans women compared to trans men. Trans women experience both transphobia and misogyny, creating a double burden of discrimination, which can eventually lead to a reduced disposition among the public to use correct gender pronouns, either through ignorance or deliberate disregard.

For tweets containing mislabelling (MISLABEL), where terms like “man” or “dude” are used to describe Individual 2, there is a significant bias towards negative sentiment. From these tweets, 33 are annotated as negative, with only 1 classified as neutral and none as positive. This indicates that mislabelling often conveys a derogatory tone. Similarly, tweets with mispronouning (MISPRONOUN), which use male pronouns like “he”, “his” or “him” to refer to Individual 2, exhibit an overwhelmingly negative sentiment. All 70 tweets with mispronouning are annotated as negative, without any positive or neutral classification.

Again, to assess the statistical validity of the data presented for Individual 2 in Table 7, a Chi-Square test was conducted, resulting in a p-value < 0.05. This indicates that the differences observed in the data are statistically significant and unlikely to have occurred by random chance.

To summarise, the data reveal a clear correlation between misgendering and negative sentiment. Among the total 200 tweets analysed, 153 are negative, with a significant portion (70) involving mispronouning and 33 containing mislabelling. The data indicate that misgendering language is often associated with derogatory or hostile contexts, emphasising the need to explore the underlying reasons behind this pattern and its implications for Individual 2 and other TGNC individuals in social media discourse.

6.2. Analysis of the Automatic Annotation

In this section, a comprehensive examination of the annotation results is conducted to uncover annotation issues. To achieve this, the section is subdivided into two parts based on the forms of annotation performed for this corpus sample: manual and automatic. An analysis of these annotations and an improvement of their guidelines is explored for further research. Furthermore, the section also covers an in-depth analysis of sentiment polarity and misgender trends observed in the corpus sample.

6.2.1. Automatic Annotation Issues

In the following section, the Python library flairNLP v0.13.1 (Akbik et al. 2019) is used to conduct sentiment analysis using deep learning models. Flair is a user-friendly NLP framework that offers pre-trained models for various tasks, including sentiment analysis. Despite its robustness, discrepancies between automated systems and human experts are common. These inconsistencies can be due to several factors, and understanding their root causes is crucial for enhancing the reliability of automated sentiment analysis (Birjali et al. 2021; Kozareva et al. 2007; Wankhade et al. 2022) in practical applications.

Consequently, this analysis delves into instances where discrepancies occur, aiming to identify their underlying causes and offer recommendations for improving the precision of automated sentiment analysis. In the context of automatic annotation, the causes of discrepancies appear to be the same for both individuals, indicating that gender does not influence these outcomes.

6.2.2. Causes of Automatic Annotation Issues

When comparing the manual sentiment annotation, where both human annotators reached a consensus, and the sentiment annotation by the flairNLP automated system, a significant disagreement was encountered. The analysis revealed that 163 tweets out of 400 had differing sentiment annotations. The causes for these differences are the following:

• Negations and double negatives: Firstly, automated sentiment detection systems, like flairNLP, can struggle with interpreting negations accurately, leading to misclassification of sentiment. This can be observed in the context of the tweet “@user1 @user2 Trans men have always been men, Individual 1 has never been a woman and is a man” where the tweet was annotated as positive by manual annotators and negative by the automatic system. In this tweet, flairNLP might have focused on the sentence “never been a woman” interpreting the negation as an indication of denying, ultimately annotating it as negative. This misinterpretation can occur because automated systems often rely on negations to comprehend the message without fully understanding the surrounding context or the deeper message conveyed by the text.
  Manual annotators, on the other hand, can recognise that the tweet is reinforcing the identity of trans men and supporting the proper use of pronouns, and the contextual understanding allows them to recognise the intended sentiment as positive, despite the presence of negations.
• Confusion between Subject and Object: Another notable challenge observed in this study is the system’s inability to distinguish between the subject and the addressee of the tweets analysed. In the tweet “Given your insistence on being a horrible person, it’s clear that understanding the basic concept that he’s a man is challenging for you. If that’s the case, then it’s best not to discuss Individual 1 at all”, the system might have interpreted “horrible person” as directed towards Individual 1, leading to a negative annotation. Although the sentiment detection system can classify the overall sentiment correctly based on the semantics of the words, it is not able to discern the direction of the comment or the intended target of criticism. Without the broader context, the system might take the phrase as literal, assuming that it is condemning Individual 1, rather than understanding that it is addressing someone who is misrepresenting or disrespecting him. This underscores the necessity for advanced linguistic models that can comprehend the context and recognise the relationships between different entities in discourse.
  Manual annotators, by contrast, can evaluate the context and correctly interpret that the term “horrible person” is directed toward someone disrespecting Individual 1. This understanding allowed them to see that the tweet’s sentiment is, in fact, positive, as it defends Individual 1’s identity and advocates for respect.
• Difficulty recognising Sarcasm and Irony: Additionally, automated systems frequently struggle with detecting irony and sarcasm, often leading to misinterpretations in sentiment analysis. For example, in the tweet “Individual 1 transitioned after enduring years of trauma from sexual abuse in Hollywood during her teenage years, following a psychotic breakdown in which she self-harmed, and after experiencing an inner voice urging her to transition [...]”, the automated system read this statement as a literal explanation for someone’s transition. However, human annotators detected the sarcasm inherent in this comment, understanding that it is questioning or mocking the notion of an “inner voice” leading someone to become trans. As a result, manual annotators classified this as a negative due to the sarcastic undertones, and flairNLP labelled it positive.
• Keyword-based analysis: The last and most prominent cause for discrepancy when employing automated sentiment analysis systems to annotate a corpus is the reliance on keyword-based analysis to classify the sentiment. This approach examines specific words and phrases to determine whether the sentiment is positive, negative or neutral. While this method can be effective for simple cases, it often fails to capture the broader context, emotional subtleties or implicit meanings that human annotators can discern. As a result, discrepancies between human annotators and these systems may arise.
  For example, the tweet “@user3 (Individual 1’s deadname) was a talented, inspiring and beautiful young woman. Individual 1 is now a disturbing, depressed ghost of their former self.” was labelled as positive by flairNLP and negative by the manual annotators. The cause might have been that the automatic system employed keyword-based analysis, identifying words like “beautiful”, “talented” and “inspiring” as indicators of positive sentiment. As a result of this focus, the system denied the derogatory use of terms such as “depressed ghost”, which eventually led to a positive annotation rather than a negative.

To summarise, automated sentiment analysis systems face significant challenges in accurately detecting sentiment, particularly in online discourse where misgendering is used against TGNC individuals. The present analysis shows discrepancies between automated and manual annotations, which reveal limitations in interpreting negations, discerning subject-object, recognising sarcasm and irony and the reliance on keyword-based analysis. Addressing these issues is imperative to improve the accuracy and reliability of automated sentiment analysis systems such as flairNLP, for future advancements in research.

6.2.3. Wordlist Frequencies

Lastly, to further explore what other forms of discriminatory language co-occur together with misgendering, comparative frequency analyses were performed using the Sketch Engine tool. It becomes possible to identify the positive or negative terms that co-occur with misgendering and their connotations by compiling a list of all unique words in the corpus sample and analysing their occurrences together with their sentiment polarity annotation. This insight is crucial for identifying which aspects related to TGNC identities receive more or less emphasis in the discourse.

To start analysing the sub-corpus targeting Individual 1, the most frequently used unique words extracted using Sketch Engine were “lesbian” (11), “tit” (9), “topless” (5), “mastectomy” (6) and “tittie” (5). To determine whether these terms were being used in a derogatory and harmful context when referring to a trans man, the following table illustrates the frequency of each term’s usage and the sentiment assigned to the tweet in which it appears (see

Table 8. Derogatory terms targeting Individual 1.

Female-Related Terms	Sentiment Polarity
	Positive	Negative	Neutral	Total
Lesbian	2	7	1	10
Tit	1	8	0	9
Mastectomy	0	6	0	6
Topless	1	3	0	4
Tittie	2	3	0	5
Total	6	27	1	34

).

Firstly, the word “lesbian” appears 11 times targeting Individual 1. Most of these instances have negative connotations, suggesting a pejorative or derogatory tone, probably to denigrate the individual’s identity or to create confusion between sexual orientation and gender identity. Other terms such as “tits” and “titties” are also frequently used in tweets annotated as negative, suggesting a tendency to objectify or feminise the individual. As for the term “mastectomy”, all tweets that include this word show negative sentiment, which may indicate a usage that focuses on gender transition surgeries as an inappropriate process. This reinforces the misperception of Individual 1’s gender.

Continuing, when analysing the list of keywords in the sub-corpus targeting Individual 2, the most frequently used unique words were “pretend” (10), “gay” (8), “dude” (6) and “manly” (3). To establish whether these terms were being used in a harmful context when referring to a trans woman, the following table illustrates the frequency of each term’s usage and the sentiment assigned to the tweet in which it appears (see

Table 9. Derogatory terms targeting Individual 2.

Male-Related Terms	Sentiment Polarity
	Positive	Negative	Neutral	Total
Pretend	2	8	0	10
Gay	0	7	1	8
Dude	0	6	0	6
Manly	0	3	0	3
Total	2	24	1	27

).

The term “pretend” is employed 10 times, 8 with a negative sentiment. This suggests an attempt to invalidate Individual 2’s gender identity by implying that she is not a woman but merely pretending to be one. The use of the term “gay”, when directed at a trans woman, may confuse the interpretation of her gender identity, equating it with sexual orientation rather than recognising her as a woman. Other terms, such as “dude” and “manly”, are explicitly masculine terms and present exclusively negative sentiment. These terms are likely to confuse Individual 2 by attributing masculine traits or identities to her, thus denying her correct gender identity.

Overall, the analysis reveals a pattern of discriminatory language strongly connected with intentional misgendering, reflecting deep biases and harmful stereotypes present within online discourse targeting TGNC individuals. Understanding these patterns is crucial for driving systemic change. This calls for a revaluation of how TGNC identities are represented on social media and the wider societal attitudes that support misgendering and discriminatory language. In addition, platforms must take responsibility for fostering inclusive and respectful discourse and implementing mechanisms to identify and address harmful language.

7. Conclusions and Future Research

In this section, a comprehensive examination of the conclusions is derived from the research questions initially posed in the introduction of this study. These research questions are addressed, focusing on understanding the linguistic phenomenon of online intentional misgendering.

7.1. Research Questions

The present study has focused on the analysis of the linguistic phenomenon of misgendering by compiling a corpus sample composed of 400 tweets addressed to two TGNC individuals, with an equal distribution of 200 tweets referring to Individual 1 and 200 referring to Individual 2 and a total of 14,492 tokens and 12,284 unique types. Subsequently, this corpus was manually annotated, and the consensual annotation between two annotators was compared with an automatic system to establish certain issues that may hinder the detection of the phenomenon. After this analysis, the questions posed at the beginning of this study are answered.

• RQ1: Does intentional misgendering as a form of microaggression perpetuate discrimination towards the TGNC community?
  The findings of this study confirm that intentional misgendering significantly perpetuates discrimination against the TGNC community. The analysis and data extracted indicate a significant correlation between intentional misgendering and negative sentiment, suggesting that misgendering indeed contributes to discrimination towards TGNC individuals. The prevalence of negative sentiment associated with misgendering underscores its role in perpetuating discrimination within online discourse.
• RQ2: Does intentional misgendering typically co-occur with other forms of aggression or discriminatory language?
  The study substantiates that intentional misgendering frequently co-occurs with other forms of aggressive or discriminatory language. The analysis reveals that intentional misgendering often accompanies other forms of discriminatory language, such as derogatory terms or negative stereotypes. The co-occurrence of misgendering with such language suggests a broader pattern of discrimination and hostility online towards TGNC individuals.
• RQ3: Is there a significant relationship between the presence of misgendering in tweets and their sentiment polarity?
  The present study’s findings strongly indicate a significant relationship between the presence of misgendering in tweets and their negative polarity. Both mispronouning (using incorrect pronouns) and mislabelling (using incorrect gender terms) consistently show a bias towards negative sentiment. This pattern of negativity is evident for both Individuals 1 and 2, indicating a consistent correlation between misgendering and negative sentiment across different contexts.
  Specifically, for Individual 1, tweets containing mispronouning predominantly exhibit negative sentiment, with a significant majority (70 tweets) annotated as negative out of 153 total negative messages. The same applies to tweets with mislabelling (33 tweets), further highlighting the correlation between misgendering and negative sentiment in this context. Similarly, for Individual 2, tweets, including mispronouning and mislabelling, lean towards negative sentiment. Thus, 78 out of 79 tweets with mispronouning were annotated as negative, and 27 out of 28 tweets with mislabelling were annotated as negatives, emphasising a strong association between misgendering and negative polarity.
  Overall, the study’s data support the conclusion that misgendering in tweets is significantly associated with negative sentiment, with 208 tweets with misgendering out of the 279 total annotated as negative. This underscores the importance of further exploration into the underlying reasons behind this correlation and its implications for TGNC individuals online.
• RQ4: Can automatic sentiment detection systems effectively identify tweets containing misgendering and expressing hatred towards transgender individuals, or is there a gap in their ability to detect this type of message?
  Automatic sentiment detection systems, such as flairNLP, face inherent limitations that result in the miscategorisation of tweets concerning their overall positivity or negativity. While these systems can sometimes correctly identify positive or negative sentiment, their broader issue lies in a contextual misunderstanding and a lack of nuance in sentiment analysis. This miscategorisation affects the system’s ability to accurately flag harmful language, including misgendering, as it struggles to correctly interpret the context in which certain words or phrases are used.
  One of the main limitations is that they rely on keyword analysis to measure the message’s sentiment. This approach often ignores the context in which positive or negative terms are used. For example, the presence of positive words in a tweet does not necessarily indicate an overall positive sentiment, as these terms may be used to refer to different persons which complicates accurate detection. In addition, these systems can not identify the addressee or subject of the sentiments leading to erroneous annotations. They operate without context concerning the individuals or groups mentioned and fail to recognise instances where seemingly harmless messages may harm others.
  Furthermore, these automated systems are inadequate at capturing the subtleties of language, including forms such as sarcasm and irony, as evidenced in this study. These linguistic nuances are often crucial in determining the true sentiment and intent of a message, but automated systems have difficulty interpreting them accurately. Hence, without the ability to grasp these subtleties, automatic systems may misclassify the sentiment of a message, leading to inaccuracies in their analysis.
  In summary, automatic sentiment detection systems face significant complications in effectively identifying tweets containing misgendering and expressions of hatred towards transgender individuals. Their reliance on keyword analysis, with a limited understanding of contextual nuances and linguistic subtleties, underscores the need for further development and refinement to enhance their accuracy in detecting and addressing this form of harmful language in online discourse.

To conclude, the present study demonstrates the prevalence of misgendering towards transgender and gender non-conforming (TGNC) individuals, particularly in the context of interactions on the social media platform X. The results reveal that intentional misgendering perpetuates discrimination towards the TGNC community and is not employed intermittently; rather, it is dominant and is accompanied by other derogatory terms that perpetuate discrimination and hostility towards this community. Hence, social media platforms must implement stricter policies and protections for TGNC individuals to foster a more inclusive online environment.

7.2. Future Lines of Research

This study calls for the implementation of robust policies by social media platforms to protect TGNC users, the development of more sophisticated natural language processing tools to better detect and address misgendering, and continued research of the linguistic and social factors contributing to this form of discrimination. Therefore, addressing these areas can create a safer and more inclusive digital environment for TGNC people, promoting their well-being and affirming their online and offline identities.

Additionally, future research could benefit from employing larger datasets to replicate and expand upon these findings. Larger samples would enable more robust analyses and enhance the generalisability of the results, providing deeper insights into the nuances of misgendering and other forms of discrimination across varied contexts.

Other future lines include the development of improved automatic sentiment detection systems which will be used for the identification of misgendering and other subtle forms of discrimination. This entails refining the corpus sample created for this study to improve contextual understanding and the ability to detect linguistic subtleties. Additionally, future research should explore how linguistic theories such as Context Theory, Inferential Pragmatics, Interactional Pragmatics, and Irony Theory can enhance the development of more sophisticated automatic detection tools. These theories can help create systems capable of understanding the complexities of language, such as sarcasm and irony, thus improving the detection and mitigation of misgendering and other forms of subtle discrimination in online interactions.

Moreover, future research should integrate insights from variational linguistics and sociolinguistics to further refine detection systems. Examining how language varies across different social groups, regions and contexts, offers valuable perspectives on misgendering and other subtle forms of discrimination. By incorporating these insights, automatic detection systems can be adapted to recognise and address diverse linguistic expressions of discrimination more effectively. This approach will enhance the ability of detection tools to operate in varied sociolinguistic contexts, leading to more accurate and contextually aware systems. Ultimately, this could contribute to creating safer and more inclusive digital environments for TGNC individuals, acknowledging and addressing the complexities of language-based discrimination in a more comprehensive manner.