Detecting East Asian Prejudice on Social Media

1 Introduction

• COVID-19 prompted concerns about East Asian prejudice and abuse, while most academic responses had focused on health and economic impacts.The introduction calls for research into other harms emerging during the pandemic.Source: §1 · p. 1
• Social media is a key setting for addressing pandemic-era social hazards, and computational models can support moderation and research on prejudice.The authors describe these tools as ways to detect and measure prejudice at scale.Source: §1 · p. 1
• The classifier distinguishes hostility, criticism, meta-discussions of East Asian prejudice, and neutral posts, achieving an F1 score of 0.83.The paper also provides a 20,000-tweet training dataset, an annotation codebook, and two analyses of associated hashtags.Source: §1 · p. 1

2 Literature Review

• East Asian prejudice has a long history, and pre-pandemic surveys and reports documented unfavorable attitudes, threats, violence, and discrimination.The review notes that official hate-crime statistics were not broken down by type of racist attack.Source: §2 · pp. 1–2
• Reports describe pandemic-era increases in Sinophobic language and anti-Chinese hashtags, alongside links between prejudice and COVID-19 misinformation.A cited analysis reported a 300% increase in hashtags supporting or encouraging violence against China during one week in March 2020.Source: §2 · p. 2
• Classifiers and training datasets specifically for East Asian prejudice were unavailable, leaving researchers to use slur searches that miss subtle prejudice and misclassify keyword-containing non-hateful content.Other datasets on racism, Islamophobia, or hate generally were not readily repurposable for this target.Source: §2 · p. 2
• Existing hate-speech tools and datasets can encode biases, misclassify language, or be fooled by spelling changes, potentially reinforcing social injustice.The authors argue development requires dialogue between machine learning and social science.Source: §2 · p. 2

3 Dataset Collection

• The final database comprised 159,320 unique English tweets collected across March 2020 and January–March 2020 sources.Initial collection yielded 96,283 unique English tweets; the additional period yielded 63,037.Source: §3 · p. 2
• Three annotators assessed the 1,000 most-used hashtags for East Asian relevance, entity, stance, coronavirus relevance, and sentiment; 97 were judged negative or very negative toward East Asia by at least one.All three hashtag annotations were made available in the repository.Source: §3 · pp. 2–3
• Researchers sampled 10,000 tweets randomly and 10,000 using one of 97 anti-East Asia hashtags to increase the likelihood of identifying prejudice.The hashtag-enriched sample was intended to make the dataset suitable for classifier training.Source: §3 · p. 3

3.1 Data pre-processing

• Hashtags were replaced with thematic labels to preserve topic cues while avoiding undue bias from their stance or sentiment.The five replacement categories distinguished East Asia, virus, East Asia-virus, other-country-virus, and other hashtags; unlisted hashtags became #HASHTAG.Source: §3.1 · p. 3
• The preprocessing addressed a trade-off: removing hashtags can obscure meaning, while retaining them can make annotators rely on particular hostile hashtags.The authors warn that such reliance can recreate a dictionary-based approach that may not generalize to contexts without those hashtags.Source: §3.1 · p. 3
• Thematic replacements preserved hashtag topics while withholding their substantive stance, and unlisted hashtags were replaced with a generic label.The paper illustrates the process by converting original hashtags into labels such as #HASHTAGVIRUS and #HASHTAGEASTASIA.Source: §3.1 · p. 3

4 Dataset Annotation

• The taxonomy and codebook were developed iteratively from theory and data because binary prejudice labels collapse distinct behaviors and can increase within-class variation.The authors sought categories that social scientists could use to generate meaningful insights.Source: §4 · p. 3

4.1 Annotation process

• Two trained annotators labeled each tweet independently, with disagreements on primary categories reviewed by an expert adjudicator before the final dataset was assembled.Two experts adjudicated cases, and the annotation effort involved 26 annotators over three weeks.Source: §4.1 · p. 3

4.2 Annotations for theme

• Annotators first judged COVID-19 and East Asia relevance; tweets not East Asian relevant were assigned to the neutral class.For theme labeling, hashtag dependence required a textual signal that the thematic replacement only confirmed, rather than a hashtag alone.Source: §4.2 · p. 4
• The detailed theme and hashtag annotation process required substantial training and helped prevent out-of-domain tweets from receiving non-neutral categories.The authors say the annotations also provide insight into how judgments were made, for research on online communications beyond prejudice.Source: §4.2 · p. 4

4.3 Primary categories

• Hostility means abuse or intense negativity toward an East Asian entity, while Criticism is negative judgment that does not cross into abuse.Negativity toward states tends to be criticism, whereas negativity toward East Asian people tends to be hostility.Source: §4.3 · p. 4
• The mutually exclusive taxonomy prioritizes entity-directed Hostility, then Criticism, Counter Speech, and Discussion of East Asian prejudice when a tweet fits multiple categories.Non-East-Asian-relevant tweets could only be Neutral; Neutral tweets could still be offensive in other ways.Source: §4.3 · p. 4

4.3.1 Annotation of Primary categories

• Agreement was moderate overall, stronger for Hostility and Neutral than for less frequent, more nuanced categories; combining Counter Speech and Discussion of East Asian prejudice yielded average Kappa 0.5.Agreement was calculated for each annotator pair, then summarized by average, minimum, and maximum.Source: §4.3.1 · p. 5
• 22% of cases went to expert adjudication; experts generally moved tweets from Neutral to other categories, especially Hostility.Neutral fell from 34% of original annotations to 29% of adjudicated decisions, while Hostility rose from 31.6% to 35%.Source: §4.3.1 · p. 5
• In 347 cases (7.7%), experts selected a category neither annotator had chosen, most often Criticism when one chose Hostility and the other None.Among these cases, Criticism rose from 18.7% of original annotations to 39.4% of expert adjudications.Source: §4.3.1 · p. 5

4.4 Secondary categories

• For Hostility and Criticism, annotators recorded targets, reconciling hundreds of initially identified variants into 78 targets across six geographic areas.Tweets could have multiple targets, with intersectional characteristics recorded where relevant.Source: §4.4 · p. 5

4.4.1 Additional flags for Hostility

• Hostility tweets received flags for interpersonal abuse, threatening language, and dehumanization.Interpersonal abuse targets an individual regardless of whether that person is East Asian; threats include willingness or advocacy of harm; dehumanization requires literal malicious comparison to nonhumans or sub-humans.Source: §4.4.1 · p. 5
• Experts did not adjudicate the secondary hostility flags, and some hostile cases had only one flag annotation or none available.When experts changed a tweet to Hostile, flags could be unavailable if neither original annotator had selected Hostility, or only one annotation could remain if the other annotator had chosen another category.Source: §4.4.1 · p. 5

4.4.2 East Asian slurs and pejoratives

• Slurs were pejorative collective terms or derivatives, while pejoratives did not need to be collective nouns; both were entered as free text.Source: §4.4.2 · p. 6

4.5 Dataset availability

• The Annotations Dataset contains all 40,000 annotator annotations and is recommended for agreement investigation and secondary-category analysis.Targets are cleaned rather than retained as free-text entries.Source: §4.5 · p. 6
• The Final Dataset contains 20,000 entries, including expert adjudications, and is recommended for training East Asian prejudice classifiers.Source: §4.5 · p. 6

• The classification task distinguishes Hostility, Criticism, Discussion of East Asian Prejudice, and Neutral, combining Counter Speech with Discussion because of low prevalence and conceptual similarity.Source: §5 · p. 6
• Models used a stratified 80/10/10 split and shared preprocessing, while contextual embeddings were compared with a one-hot LSTM baseline.Preprocessing removed URLs and usernames, lowercased text, and replaced hashtags with generic or thematic tokens.Source: §5 · p. 6
• RoBERTa achieved the highest F1 score (0.83), a 7-point improvement over the LSTM baseline (0.76).Source: §5 · p. 6, Table 4 · p. 7
• The confusion matrix shows strongest performance for Neutral (91.6% recall, 93% precision) and more errors between the similar Criticism and Hostility categories.Source: §5 · pp. 6–7, Figure 1 · p. 7

6.1 Annotator Errors, (17% of total)

• About 17% (58) of reviewed errors were cases where the model’s prediction may have fit the tweet and taxonomy better than the annotation; this is not the dataset-wide misannotation rate.The sample was selected because the model made an error, so actual misannotation prevalence is likely far lower.Source: §6.1 · p. 7
• Of 58 annotator errors, 36 were clear primary-category misapplications; 22 involved detailed annotations on irrelevant tweets, often after over-reliance on hashtags.An early misannotation could lead to flawed subsequent annotations.Source: §6.1 · p. 7

6.2 Machine Learning Errors, (83% of total)

• Of total errors, 83% were due to model errors: 37% were non-edge cases easily identified by humans, and 46% were ambiguous edge cases where misclassification had some merit.Source: §6.2 · p. 7

6.2.1 Edge Cases

• Edge cases included Hostility–Criticism confusion, Neutral tweets mistaken for prejudice discussions, co-present categories forced into one label, and ambiguous wording.Normatively framed criticism could be mistaken for Criticism rather than Hostility; tweets discussing the virus without explicit prejudice could be mistaken for prejudice discussions.Source: §6.2.1 · pp. 7–8

6.2.2 Non-edge Cases

• Non-edge mistakes included over-sensitivity to category-associated phrases, wrong-target classifications, missed sarcasm or innuendo, and failures requiring cultural context.More than a third of non-edge errors involved an appropriate category applied to a non-East-Asian target.Source: §6.2.2 · p. 8

6.3 Addressing Classification Errors

• Better annotation processes and guidelines may limit avoidable errors, while larger, more balanced datasets may reduce edge cases and larger datasets or advanced architectures may address obvious model mistakes.The authors note that eliminating errors entirely is arguably impossible and that edge cases have no easy solution.Source: §6.3 · p. 8

7 Hashtag analysis

• The analysis retained hashtags used in at least ten tweets per category and ranked them by the share of uses accounted for by that category; only the twenty strongest Hostility associations are shown.Source: §7 · p. 8
• Some hashtags were highly likely to appear only in tweets expressing Hostility against East Asian entities, but #fuckchina and #blamechina were hostile in only 67.5% and 60.5% of their uses.The authors argue that this variability makes a purpose-built classifier preferable to relying on hashtags and keywords alone.Source: §7 · p. 8
• Table 5 is identified as listing hashtags in hostile tweets.Source: Table 5 · p. 9

8 Conclusion

• The authors provide a classifier for detecting East Asian prejudice.Source: §8 · p. 10
• The repository includes a 20,000-tweet training dataset and a separate dataset containing 40,000 annotations made by individual annotators.Source: §8 · p. 10
• The materials include a list replacing COVID-19-specific hashtags with thematic replacements.Source: §8 · p. 10
• The repository includes three annotation sets for the 1,000 most-used hashtags, covering COVID-19 relevance, East Asian relevance, and stance.Source: §8 · p. 10
• The full codebook provides hate-speech annotation guidelines, information, and examples.Source: §8 · p. 10