Latent Hatred Taxonomy Overview

• Latent hatred taxonomy is a system that defines and classifies covert hate speech through semantic, rhetorical, and structural strategies.
• It employs codetype, six-fold, modular, and universal frameworks to systematically capture and annotate implicit hate signals in multiple languages.
• Integrating these taxonomies with computational models enhances detection accuracy, supports reproducible methods, and improves moderation practices.

Latent hatred taxonomy refers to fine-grained systems for classifying implicit (covert, coded, or indirect) hate speech, distinguishing it from explicit hate expressions and enabling more robust computational detection. Unlike surface-level approaches that rely on observable slurs or threats, latent hatred taxonomies systematically capture semantic, rhetorical, or structural strategies by which prejudicial attitudes are encoded in seemingly innocuous language. Recent taxonomic frameworks—most notably codetype-based systems and conceptual element decompositions—have demonstrated substantial impact on both annotation practice and LLM detection performance across multiple languages and platforms.

1. Core Taxonomies for Implicit/Latent Hate

Two primary lines of work provide formal taxonomies for latent hatred, distinguishing specific encoding strategies or conceptual facets:

Codetype Taxonomy

"Cracking the Code: Enhancing Implicit Hate Speech Detection through Coding Classification" introduces a six-way codetype taxonomy, where a codetype is a "hate speech encoding strategy" modulating non-hateful surface forms to covertly signal prejudice. After annotation on the Chinese ToxiCN corpus and ISHate English data, six dominant codetypes were identified (Wei et al., 5 Jun 2025):

Codetype	Formal Label	Strategy Description
Abbreviation	$c_1$	Shortened form (acronym/pinyin) masking prejudicial references
Metaphor	$c_2$	Figurative mapping, e.g., animalistic or disease metaphors
Suggestion	$c_3$	Implicit recommendations toward discrimination or exclusion
Comparison	$c_4$	Analogizing targets to negative classes/invoking hierarchical schemas
Irony	$c_5$	Sarcastic modality, often the negation of literal meaning
Misinformation	$c_6$	Factually false claims constructed for negative attributions

A residual "Other" category captures hybrid or rare codings.

Six-Fold Implicit Hate Speech Taxonomy

ElSherief et al.'s Latent Hatred taxonomy (ElSherief et al., 2021) is similarly six-fold, anchored in social scientific literature:

• White Grievance: Rhetoric reframing majority groups as "victims" of reverse discrimination.
• Incitement to Violence: Calls (explicit, implicit) for violence or group-based power assertion.
• Inferiority Language: Dehumanization or toxification (insect/disease metaphors).
• Stereotype Endorsement: Repetition or affirmation of negative group stereotypes.
• Animosity: Expressions coded to evoke hostility but lacking direct attack.
• Exclusion: Implicit endorsement of exclusion from resources, benefits, or society.

Empirical analysis shows these axes collectively cover 98.6% of annotated implicit hate in U.S. hate-group Twitter timelines.

2. Modular and Universal Taxonomy Frameworks

To address the lack of interoperability and the underrepresentation of latent forms in prior taxonomies, recent work proposes modular and universal frameworks:

Conceptual Element Modular Taxonomy

"A Modular Taxonomy for Hate Speech Definitions and Its Impact on Zero-Shot LLM Classification Performance" decomposes the notion of hate speech into 14 Conceptual Elements (CEs) (Melis et al., 23 Jun 2025). This taxonomy includes both explicit and latent dimensions:

• Foundational: Form of Communication (FoC), Target (T), Problematic Content (PC), Addressed Attributes (AA)
• Extensive-of-Foundational: E.g., elaborated definitions for each foundational element (EDFoC, EDT, EDPC, LAA)
• Accessory: Social and individual implications (sPI, iPI), Exceptions, Implicit Hate Speech (IHS), Examples, Reference to Law

The IHS element encompasses "irony, stereotypes, or misinformation," closely aligned with codetypes in (Wei et al., 5 Jun 2025). Definitions are algorithmically constructed by selecting subsets of CEs with an invariant preamble and string concatenation.

Universal Hierarchical Taxonomy

"Improving Hate Speech Classification with Cross-Taxonomy Dataset Integration" formalizes a universal, latent taxonomy $U$ that subsumes heterogeneous datasets (Fillies et al., 7 Mar 2025). At the top level:

• No-Hate
• Hate: further decomposed as
  • Target_of_hate (5 branches, 43 leaves—e.g., Religion, Race_Ethnicity, Physical_attributes)
  • Types_of_hate (Derogation, Animosity, Threatening_Language, Support_for_Hateful_Entities, Dehumanization)

A set of deterministic mapping functions $f_i : L_i \to \{0,1\}^{|U|}$ aligns original dataset labels to the universal space, supporting robust, multi-label classification.

3. Annotation and Corpus Construction Methodologies

The creation of latent hatred taxonomies typically involves iterative corpus analysis, inductive coding, and linguistic or social-scientific theory integration. Key methodologies across works include:

• Selective Subsampling and Manual Annotation: Codetypes in (Wei et al., 5 Jun 2025) were induced by annotating filtered ToxiCN/ISHate data, with annotator agreement ensured via example-rich guidelines.
• Inductive Conceptual Coding: (Melis et al., 23 Jun 2025) harvested 20 definitions from legal, policy, and academic sources, abstracted atomic CEs, and organized these into layered categories via style-controlled snippet templates.
• Cross-Taxonomy Mapping and Correction: (Fillies et al., 7 Mar 2025) aligned overlapping and semantically close labels via word-level mapping and error analysis; integration cycles corrected misaligned classes (e.g., mapping "non_white" as an ancestor of "Black").

These processes ensure maximized coverage of real-world phenomena while supporting reproducibility and interoperability.

4. Integration With Computational Models

Latent hatred taxonomies are increasingly leveraged as schema for LLM prompt engineering, embedding, and multi-label classification:

• Prompt-Based Integration: (Wei et al., 5 Jun 2025) implements two routes—prompting LLMs to classify sentences by codetype; and embedding codetype information directly for context conditioning. Explicitly incorporating codetype boosted macro-F1 on both Chinese and English tasks.
• Zero-Shot + Modular Definitions: In (Melis et al., 23 Jun 2025), prompts with varying CE combinations are evaluated with Llama-3-8B, Mistral-7B, and Flan-T5-XL on synthetic, human-curated, and real-world test sets. Detailed definitions improve recall (reducing false negatives) for implicit/latent signals, especially when including IHS.
• Multi-Label Hierarchical Classification: (Fillies et al., 7 Mar 2025) fine-tuned RoBERTa on the universal taxonomy target, with per-class sigmoid outputs and binary cross-entropy loss, yielding F1 of 0.84 on an independent YouTube test set after taxonomy correction and mixed human–machine annotation cycles.

Architecture-specific response to definition granularity is prominent: Mistral-7B is highly sensitive to Accessory CEs (notably LAA and IHS), while Flan-T5-XL robustly benefits from explicit, extended definition blocks (Melis et al., 23 Jun 2025).

5. Impact, Evaluation, and Practical Consequences

Fidelity to latent hatred taxonomies markedly enhances both research and operational detection of subtle hate phenomena:

• Empirical Gains: Augmenting LLM prompts with codetypes or modular conceptual elements improves detection rates of implicit hate, especially among under-represented forms (e.g., dog whistles, metaphors, irony) (Wei et al., 5 Jun 2025, Melis et al., 23 Jun 2025).
• Coverage: Detailed frameworks (six-fold, codetype, modular) account for >98% of implicitly hateful content in benchmark corpora (ElSherief et al., 2021).
• Adaptivity: Modular and hierarchical approaches allow platforms to "dial in" recall versus precision by selecting appropriate taxonomy elements, facilitating context-specific moderation policies (Melis et al., 23 Jun 2025, Fillies et al., 7 Mar 2025).
• Interoperability: Universal and modular taxonomies enable cross-resource and federated learning, reducing dependence on dataset-specific models and improving generalizability (Fillies et al., 7 Mar 2025).
• Limitations: Overly detailed definitions may increase false positives in benign discourse (notably for conservative models like LLama-3), and partial refusals occur for longer prompt compositions (Melis et al., 23 Jun 2025).

6. Open Challenges and Future Directions

Unresolved issues and ongoing research avenues include:

• Automatic Taxonomy Refinement: Symbolic ontology-matching and gradient-based CE selection to streamline taxonomy adaptation (Fillies et al., 7 Mar 2025, Melis et al., 23 Jun 2025).
• Latent Embedding of Taxonomic Structures: Learning continuous latent representations for taxonomy nodes to capture non-binary semantic proximity.
• Extensibility to Related Constructs: Modular frameworks may extend to adjacent domains (misinformation, mental-health discourse) to disentangle ambiguous harm dimensions (Melis et al., 23 Jun 2025).
• Reliability and Bias Auditing: Iterative annotation, error analysis, and supervised correction cycles are necessary to address annotation ambiguity and subclass misalignment.

The rapidly expanding literature on latent hatred taxonomy demonstrates both the theoretical necessity and empirical utility of structure-aware classification frameworks for implicit hate, confirming their centrality in modern NLP moderation and analysis pipelines.