Introduction

Over the last few years, news coverage of transgender (trans) people has increased dramatically. While some articles were published in the interest of educating the mainstream public about gender-diverse people, much of the coverage involves policies and court rulings that negate the existence or restrict the public presence of trans people. Within 100 days of returning to the White House, the Trump administration barred members of the press from the Oval Office for refusing to adopt a new name for the Gulf of Mexico.1 The administration has already issued executive orders to stop health care providers from treating trans or nonbinary minors, ban trans people from serving in branches of the military, and to move trans women incarcerated in federal prisons to men’s prisons.2

The news media sets the official, public record of events that generate public discourse. There is clear incentive for the mainstream news media, regardless of political ideology, to conform to narratives most favorable to the Trump administration. In that sense, social, particularly gender, norms may be one of the simplest ways for members of the press to gain access to a calamitous administration. Given the Trump administration’s unambiguous rejection of the existence of transgender people, it stands to reason that media sources with an ideological bias that aligns with the current administration would use more negative rhetoric in news articles about transgender people than news sources without the same particular bias. In the days following the second presidential inauguration of Donald Trump, do news sources with right-wing media bias ratings use more negative language to depict or discuss transgender people than news sources with other political bias ratings? To answer this question, a careful sentiment analysis of news about trans people must be performed. One approach is to determine what topics are being associated with the trans community. This provides some context for the objectified sentiment valence: what issues are generated emotionally charged language in the media? From there, the average sentiment for each political bias rating group determines whether there is any real difference in the language used.

Literature Review

The mainstream media sets the agenda for public opinion on emerging issues, especially those in marginalized communities. In “Writing in the Margins: Mainstream News Media Representations of Transgenderism”, Thomas Billard summarmises why representation of trans people in the news matters: news media is the main vector by which marginalized communities legitimize their identities and the political issues affecting them in public discourse.3 As trans people become more visible in society, there are more news articles and public discourse questioning where trans people are allowed to be visible. Pew Research Center survey results from February 2025 estimated that two-thirds of respondents support laws that would “require trans athletes to compete on teams that match their sex assigned at birth”. In contrast, although more than 55% of respondents support laws that would ban health care professionals from treating minors who wanted to transition, the same number of respondents support laws that would protect trans people from housing or employment discrimination.4 These results suggest that there is some general consensus that trans people are a legitimate marginalized community. However, in public discourse, trans identities are not so legitimated (or, well understood) that they are allowed to move unrestricted in society. This distrust stems from stereotypes of trans people as deceptive, predatory and possibly sexually deviant.5

Since the volume of news articles about trans people has increased the most within the last decade, there are few standard coding schemes for news articles about that specific community. Olveira-Araujo created a list of indicators used to hand-code documents for variables such as misgendering and deadnaming. The most relevant of those indicators is “wrong body discourse” which is defined as “represent[ing trans experiences] as an incongruence between sexual identity and the body.”6 News articles with a right-wing media bias may use language implying that trans people are at odds with their bodies or human biology in an effort to de-legitimize or erase trans people’s lived experiences. The common assumption is that news articles from politically right-wing sources use more negative words in relation to transgender people than articles from centrist or leftist sources.

Data, Methods & Analysis

Data

The dataset was collected from NewsAPI.org using a keyword search on the word “transgender”. The API limits users’ calls to a small number of responses per day, so news articles were retrieved in batches according to news source and availability. The document corpus includes 650 news articles published betweem March 27 and May 6 from the following news publications: Fox News, Breitbart News, ABC News, Newsweek, BBC News, CBS News, Forbes, CNN, NPR, The Atlantic, Deadline, Rolling Stone, ProPublica, USAToday, Vox, Democracy Now!, and The Hill. The corpus uses the descriptions for the articles to reduce processing runtime.

Only news articles from sources with an AllSides ideological media bias rating were included in the dataset. AllSides Media Bias Ratings are a five-point system (Left, Lean Left, Center, Lean Right and Right) based on editorial panels, surveys and community feedback.7 The original scores have an uneven distribution with over 25% of articles assigned Lean Left and 0% of articles assigned to Lean Right. A modified three-point version of the Allsides Media Bias Ratings (Left, Center & Right) was used to provide a more balanced bimodal distribution of articles between centrist, leftist and right-wing sources, representing 25%, 34% and 40% of the documents respectively.

Methods

The analysis for this project includes LDA topic modeling and sentiment analysis using the AFINN lexicon. LDA topic modeling identifies the terms most closely associated with topics related to transgender people. This modeling is done to ensure that publications are not covering completely different events altogether. After identifying the words most associated with trans news for each political bias group, a dictionary-method sentiment analysis can be performed to determine if right-wing sources use more negative language in news articles about trans people.

Topic Modeling

Articles from mainstream sources are likely to cover many of the same political and cultural news. When comparing news articles across political bias ratings, the articles can be categorized by topic (or issues) to confirm which events have entered the news media. The Latent Dirichlet Allocation (LDA) topic model is an unsupervised method for discovering terms for each topic within a document corpus. While this is a useful method for exploring data, there is no guarantee that the topics or the terms used to define them will be meaningful. Korenčić et al. used a more sophisticated method to measure the media agenda using topic modeling to identify themes covered in the media.8 This project is simplified in that the high-level theme has already been determined: transgender people. In this case, the purpose of the topic modeling is to identify the issues affecting the trans community that have entered public discourse and to determine that no political bias group is alone in reporting on any of these issues.

Sentiment Analysis

After confirming that these publications generally cover the same events across the political spectrum, sentiment analysis can be used to determine if there is any significant, emotionally charged difference in the language used in stories related to trans people. The AFINN popular lexicon features a dictionary of several thousand words with sentiment values (valences) on a 10-point scale from negative (-5) to positive (5) with 0 representing neutrality.9 Using the dictionary method to assign sentiment valence, the average sentiment for news articles can be calculated for each political bias group. Although dictionary methods for sentiment analysis are not the most valid or accurate10, this method was chosen for being less time-consuming and requiring fewer people than hand-coding documents.

Analysis

Is there any difference in the topics covered between political ratings?

Figure 1 shows the 10 terms most closely associated with each topic. The topic model for this corpus produced 7 topics roughly related to the following events in the news: girl’s sports teams banning trans athletes; a trans rights bill in Colorado; an executive order banning trans people from the military; a cis fencer refusing to compete against a trans opponent; the Department of Justice suing the Maine Department of Education for not banning trans athletes from sports teams; and a miscellaneous category that broadly mentions the LGBTQ community. For topic 1, the headline for one of the top documents reads “Bondi Announces DOJ Lawsuit Against Maine over Trans Sports Dispute”; for topic 3, one of the headlines is “Trump asks Supreme Court to lift block on transgender military ban.” Note that topics 1 and 6 may be overlapping topics referencing the same events: legal controversy around trans women and girls playing on sports teams with cis women and girls. One of the benefits of LDA modeling is that documents are treated as a mixture of topics, meaning that any document can contain elements of different topics.

Here the documents are labeled by topic according to their maximum LDA model gamma results. For the first 2 topics, articles from centrist publications appear to be less represented. In fact, topic 1 appears to be dominated by articles from right-wing publications. For topics 3 through 7, there seems to be a roughly equal number of documents from different bias groups in each topic.

Table 1. ANOVA for effect of topic and bias rating on topic-bias document counts

##                Df Sum Sq Mean Sq F value Pr(>F)
## political_bias  2  737.2   368.6   1.827  0.203
## topic           6  267.0    44.5   0.221  0.963
## Residuals      12 2420.8   201.7

When taking the differences between assigned topics and political bias ratings into account, neither variable have a significant impact on the documents assigned to each topic. The results of an ANOVA test are not significant for either variable in relation to the number of documents in each topic. This suggests that, despite the slight imbalance in topic 1, no news topic is disproportionately covered by any one political bias group.

As stated earlier, a topic derived from an LDA model does not have a one-to-one relationship to any one document. In this case, for cross-validation, the LDA model topic assignments are the predicted values compared to documents that actually contain those terms as the reference values. Reference values for a topic are determined by searching for the topic terms in each document and assigning to the document the topic that has the most terms matched. Figure 2 shows how topic 1 (trans athletes in girl’s sports) is the most recognized topic in both the model and the texts themselves. Over 25% of the documents in the corpus have words that appear in the key terms for topic 1. According to this confusion matrix, the overall accuracy of the model is 0.20. This is supported by the low number of correct predictions for almost all of topics. For the other topics, the share of documents ranges from 6 to 17%.

Is there any difference in sentiment between political ratings?

Table 2. ANOVA for effect of topic and bias rating on average sentiment

##                 Df Sum Sq Mean Sq F value Pr(>F)
## political_bias   2    5.5   2.749   1.123  0.326
## Residuals      495 1211.6   2.448

The boxplot in Figure 3 and the ANOVA results in Table 2 show the same conclusion: the political bias rating has no significant impact on sentiment in news articles. Although the difference is not significant, it is worth noting that in Figure 3 the median sentiment from right-wing sources is around 0.5 valence points lower (more negative) than the median sentiments from leftist and centrist sources. Meanwhile the sentiment distributions for leftist and centrist publications are more similar, especially around their medians and third quartiles.

The t-SNE plot re-contextualizes multiple dimensions into a number of dimensions that can be represented graphically. The significance of the plot is not in the cluster sizes or the distances between them.11 This plot provides a 2-dimensional map that shows, across media bias groups, how frequently certain words appeared and their sentiment valence. The data for this plot is based on the word embeddings for the words in the AFINN lexicon, the only words that are relevant for this dictionary-method of sentiment analysis. The word embeddings provide context for words within a document by determining each word’s relationship and similarities with other words.12

Across the political bias plots in Figure 4, several words overlap across media bias groups such as “ban”, “protest” and “allow”. Only the right-wing media publications include a word with an especially high-valence score: “win”. Although these results are not significant, the presence of words like “refuse”, “fire”, and “punish” may explain the lower median sentiment valence for right-wing publications. Note that the word “ban” is also one of the terms associated with topic 1 from the LDA model. The presence of words like “win” and “punish” imply a sense of victory and domination in response to recent events. Incidentally, the majority of the words that were embedded and appeared frequently enough may not have been as relevant to sentiment analysis as to topic modeling. Interestingly, in the context of trans issues, words like “biological” might have a more negative sentiment attached given how it is used to delegitimize the lived experiences of trans people.

Conclusions & Limitations

A simple topic modeling and sentiment analysis process suggests that news articles from right-wing publications do not use significantly more negative language when discussing trans issues than other publications with known political bias leanings. From late March to early May, the public discourse around trans people revolved around trans women’s participation in women’s sports. In fact, two other topics identified by the unsupervised model seem to overlap with a general theme of trans women being banned from women’s sports. Although there is a small difference in average sentiment valence, the difference is not enough to be considered statistically significant. With the help of the t-SNE plot, the relative frequency of words like “ban” and “protest” seem to drive the negative sentiment across all political bias ratings.

This question could be further explored using a more robust corpus of news articles with an even or normal distribution of political ratings. Overall, the model could benefit from a larger dataset with more centrist news sources and more variety among right-wing publications. While the sources include a variety of publications with different political biases, the corpus would benefit from including news publications with wider readership such as the New York Times or the Wall Street Journal. A longer time frame would also provide more information on how trans issues have developed over time and how they have been introduced into public discourse. Even a comparison of the last five years would show more about how trans issues have developed to the point of becoming a salient political issue resulting in overreaching executive orders. In this case, the corpus was limited to the data available from the News API.

Unsupervised machine learning methods were used for this analysis. Supervised analysis, with the benefit of hand-coded, gold-standard validation data sets, could produce more valid results. Indicators for news media that delegitimize the existence of trans people need sophisticated coding schemes that, at this point, can only be reliably executed by hand by coders knowledgeable about diverse genders and the trans community. As previously noted, certain words and phrases, like appeals to “biology”, may appear inoffensive to a wider audience but can in fact be delegitimizing of the existence of trans people.

References

  1. (Bauder, 2025)↩︎

  2. (Mulvihill, 2025)↩︎

  3. (Billard, 2018)↩︎

  4. (Pew Research Center 2025)↩︎

  5. (Billard, 2018)↩︎

  6. (Olveira-Araujo, 2024)↩︎

  7. (AllSides2025)↩︎

  8. (Korenčić et al., 2015)↩︎

  9. (Quanteda, 2024)↩︎

  10. (van Atteveldt et al., 2021)↩︎

  11. Wattenberg, 2016↩︎

  12. (Saito, 2019)↩︎