HatefulDiscussions Dataset

• HatefulDiscussions Dataset is a curated collection of complete social media conversation threads focused on Jew-directed hate speech and counter responses.
• It employs a dual-dimensional, multi-label annotation scheme capturing both thematic nuances and rhetorical appeals (Logos, Ethos, Pathos) within each tweet.
• The dataset comprises 92 threads with 720 tweets from X over a five-year period, enabling fine-grained qualitative and quantitative analysis of online hate dynamics.

The HatefulDiscussions Dataset is a curated collection of annotated social media conversations sourced from the X platform (formerly Twitter), designed to enable systematic study of the simultaneous emergence and contestation of Jew-directed hate speech and its counter-responses within naturalistic conversational contexts. The dataset, introduced in "Dialogues of Dissent: Thematic and Rhetorical Dimensions of Hate and Counter-Hate Speech in Social Media Conversations" (Levi et al., 28 Jul 2025), incorporates a dual-dimensional, multi-label annotation scheme capturing the thematic and rhetorical nuances of both hate and counter-hate discourse. The corpus is foundational for research on online antisemitism, conversational dynamics of hate, and mechanisms of counter-speech.

1. Dataset Composition and Collection Protocol

The HatefulDiscussions Dataset comprises 92 complete conversation threads, encompassing a total of 720 tweets (nodes), each forming a node in a conversational reply chain. The corpus is drawn from public posts on X (formerly Twitter) spanning August 26, 2018, through January 22, 2023. The collection utilized a lexicon of terms related to Jews and Judaism, including Zionist-related expressions, obtained from the lexicon prepared by Ron et al. (2023), specifically selected to maximize the recall of Jew-directed hate speech.

Tweets were collected as follows:

1. Crawl: All public tweets matching the target filter lexicon were retrieved via the Twitter API v2.
2. Conversation Reconstruction: Complete reply chains (i.e., the full conversation trees) were programmatically reconstructed.
3. Manual Selection: A subset of 92 conversation chains was manually identified for inclusion; each thread contains (a) at least one instance of hate speech directed at Jews and (b) at least one response characterized as counter-hate.
4. Post-filtering: Threads not meeting both criteria (i.e., those lacking either hate or counter-hate) were excluded.

The selection protocol ensures that each conversation models the interaction between oppositional speech acts and their responses within a naturalistic online discourse environment (Levi et al., 28 Jul 2025).

2. Annotation Scheme: Thematic and Rhetorical Dimensions

The dataset's annotation framework is multi-labeled and orthogonal along two dimensions:

• Thematic Dimension: Categorizes the discursive aspects or topical content of each message.
• Rhetorical Dimension: Captures the manner or mode of communication, operationalized via Aristotelian rhetorical appeals.

Thematic Categories

Hate Speech (any subset of the following may be present):

1. Contemptuous Characterization (contempt): Characterizations expressing intolerance, prejudice, strong dislike, or constructing the target as inferior.
2. Call for Anti-Group Action (call_for_action): Calls for harmful, or potentially harmful, actions (including violence or discrimination) against the target.
3. Undermining of Fundamental Narratives (under_narrative): Statements belittling, mocking, or denying foundational elements of the target group’s history or identity (e.g., Holocaust denial).

Counter-Hate Speech (any subset may be present):

1. Counter-Narrative (counter_narrative): Efforts to restore positive or accurate representations of the targeted group by affirming their values or revising group narratives.
2. Defensive Rebuttal (def_rebuttal): Direct responses refuting the hate speech content, correcting misinformation, or highlighting inaccuracies.
3. Call for Positive Action (call_for_pos_action): Language that promotes action or advocacy countering hate (distinct from mere rebuttal).

Rhetorical Categories

Each message—regardless of type—is further labeled with one or more of the following rhetorical appeals:

1. Logos: Use of logical arguments, evidence, and reasoning.
2. Ethos: Appeals based on the speaker’s credibility or moral authority.
3. Pathos: Evocation of emotion to persuade or influence.

Messages are jointly annotated for thematic and rhetorical properties, permitting multi-label assignment per tweet (Levi et al., 28 Jul 2025).

3. Data Structure and Representation

Each conversation in the dataset is encoded as a graph-structured tree, with tweet nodes connected via reply-relation edges. Nodes are attributed with metadata, thematic labels, and rhetorical labels. The dataset includes identifiers for conversation thread, tweet text (where permissible under Twitter's TOS), user metadata (availability not specified), and assigned labels.

A partial tabular representation:

Node ID	Thematic Labels	Rhetorical Labels
T123	contempt, call_for_action	pathos
T124	counter_narrative, def_rebuttal	logos, ethos
...	...	...

The data model supports multi-label assignment per node in both dimensions.

4. Context: Sampling, Limitations, and Applicability

The dataset is thematically constrained to discourses involving Jew-directed hate, as an explicit product of the selection lexicon and manual thread curation. The concentration on Jew- and Judaism-related terminology, together with intentional selection for threads exhibiting both hate and counter-hate, makes the corpus specialized for studying antisemitism and response dynamics rather than hate speech in general or random conversational samples.

No statistics are provided regarding inter-annotator agreement, annotation protocol details beyond category definitions, or dataset split recommendations.

A plausible implication is that, given the small sample size and high thematic focus, the dataset is optimized for fine-grained qualitative analysis, thematic-rhetorical interaction study, and pilot quantitative modeling rather than for high-capacity training of large neural architectures.

5. Analytical Dimensions and Findings Enabled

The dataset’s joint thematic-rhetorical annotations support:

• Statistical analyses of co-occurrence and transition patterns between hate and counter-hate speech types.
• Mapping of rhetorical mode prevalence across discursive stages (initiation, escalation, or resolution).
• Correlational studies on the public engagement metrics (e.g., likes, retweets) of different category combinations.

The paper (Levi et al., 28 Jul 2025) empirically analyzes the frequency and rhetorical composition of both hate and counter-hate messages, providing granular insights into the strategies employed for both dissemination of hate and counter-narrative resistance.

6. Relations to Existing Scholarship and Future Directions

The lexicon used derives from Ron et al. (2023), indicating explicit methodological lineage with prior work maximizing recall for Jew-directed hate. The dataset’s focus on conversational context and response categorization addresses calls in the literature for moving beyond single-message hate detection towards understanding hate–counter-hate interaction and contextualized discourse analysis.

Possible future expansions could include increased topical diversity, application to other target groups, annotation scale-up, and the integration of further conversational features (user stance, temporal dynamics), although these are not addressed in the source publication.

7. Summary Table: Dataset Characteristics

Dimension	Value/Description	Source
Conversations	92 complete threads	(Levi et al., 28 Jul 2025)
Tweets	720 nodes	(Levi et al., 28 Jul 2025)
Platform	X (Twitter)	(Levi et al., 28 Jul 2025)
Time Range	2018-08-26 to 2023-01-22	(Levi et al., 28 Jul 2025)
Lexicon Focus	Jew-, Judaism-related keywords, incl. Zionist references	Ron et al. (2023)
Annotation Scheme	Multi-labeled: thematic + rhetorical (Logos, Ethos, Pathos)	(Levi et al., 28 Jul 2025)
Hate/Counter-Hate Required	Yes (for inclusion)	(Levi et al., 28 Jul 2025)

This table summarizes the dataset’s scope and design features as described in the source.