C-HarmKB: Chinese Harmful Semantic Knowledge Base

• C-HarmKB is a specialized repository of Chinese cultural concepts and offensive vocabulary that underpins harmful meme attribution.
• It is constructed by expanding existing insult lexicons and harvesting data from encyclopedic and subculture sources, then manually annotating for cultural context.
• The system uses vector-based hybrid retrieval and reranking to bridge multimodal cues with cultural semantics, improving detection performance.

Searching arXiv for papers on C-HarmKB and closely related Chinese harmful-content resources. Chinese Harmful Semantic Knowledge Base, commonly abbreviated C-HarmKB, is introduced as “the first Chinese harmful semantic knowledge base” and more concretely as “a specialized knowledge base of Chinese cultural concepts and offensive vocabulary” for the interpretation of Chinese harmful memes. In the literature, it is presented not as a generic lexicon, but as a curated repository of Chinese cultural concepts, offensive vocabulary, and contextual definitions designed to provide essential prior knowledge for culturally grounded harmfulness attribution, especially where multimodal cues are ambiguous or semantically indirect (Wang et al., 23 May 2026).

1. Origin within Chinese harmful meme research

C-HarmKB emerged from a line of work arguing that Chinese harmful meme detection is not exhausted by image-text classification, because many memes depend on local internet slang, subcultural conventions, derogatory expressions, historical references, and social stereotypes. Earlier work framed the field through ToxiCN MM, described as the first Chinese harmful meme dataset with 12,000 samples and a baseline detector called Multimodal Knowledge Enhancement (MKE) (Lu et al., 2024). The later attribution-oriented framework centered on Ex-ToxiCN-MM, described as the first Chinese harmful meme explanation dataset, and introduced C-HarmKB as the knowledge substrate needed to distinguish opposing plausible readings of a meme (Wang et al., 23 May 2026).

The motivation for C-HarmKB is tied to two claims. First, “accurately assessing a meme's harmfulness depends heavily on understanding deep cultural context.” Second, “many memes are semantically ambiguous, making harmfulness highly subjective.” The paper illustrates this with expressions such as “vegetable dog” ($cài gǒu$), where literal interpretation misses the fact that “vegetable” can mean “unskilled” or “low-level” in Chinese internet slang and “dog” can function as a self-deprecating suffix, and with “Gyarados” versus “salted fish”, where the same meme can support both harmful and non-harmful interpretations (Wang et al., 23 May 2026).

This positioning distinguishes C-HarmKB from a simple profanity list. It is intended to supply the missing background semantics that models fail to recover from surface image-text signals alone. A plausible implication is that C-HarmKB addresses not only lexical offensiveness, but also the interpretive gap between multimodal evidence and culturally situated harmful intent.

2. Construction, scope, and contents

The paper states that C-HarmKB contains 2,870 entries. It was built by “expanding upon an existing Chinese insult lexicon” and by harvesting terms and definitions from Baidu Baike, Wikipedia, and specialized subculture forums using automated web crawlers. It was then “further enriched with contextual information gathered during the creation of the Ex-ToxiCN-MM dataset”, where annotators documented references to public figures, landmarks, events, or artworks in a candidate list that was later screened for KB inclusion (Wang et al., 23 May 2026).

The construction process is described as a multi-step pipeline: seed expansion from an existing Chinese insult lexicon; automated harvesting from encyclopedic and forum sources; candidate accumulation during dataset annotation; screening and filtering; and manual annotation and categorization. The paper specifies that three native Chinese speakers with expertise in internet culture categorized each term into predefined classes and authored detailed definitions explaining the term’s cultural context, origin, and harmful connotations (Wang et al., 23 May 2026).

The predefined classes explicitly named are:

• Sexism
• Racism
• Region
• LGBTQ
• Others

The described contents include Chinese internet slang, offensive expressions, cultural concepts, discriminatory language, cyberbullying-related terms, historical events, and sociocultural contexts. The paper therefore characterizes the KB as broader than a slur inventory. It includes culturally grounded concepts needed to interpret harmful semantics, including stereotype-bearing references and historically loaded terms (Wang et al., 23 May 2026).

At the same time, several details are explicitly not provided. The paper does not give a complete schema definition, relation types, inter-annotator agreement for KB annotation, per-category counts, coverage rates over Ex-ToxiCN-MM, or licensing and format details for the KB itself. This absence is consequential: C-HarmKB is presented as a curated knowledge base, but not as a fully specified ontology or graph schema.

3. Representation and retrieval architecture

C-HarmKB is represented primarily as textual knowledge entries, not as a symbolic graph. The paper states that “the textual content of the knowledge base is segmented and converted into vector representations using a pre-trained Chinese word embedding model,” and that these representations, encoding vocabulary, classifications, and definitions, are stored in a vector database (Wang et al., 23 May 2026).

This representation has two immediate implications. First, the operational KB unit is a textual fragment containing at least a vocabulary item, a class label, and a definition. Second, retrieval is formulated as hybrid lexical-semantic search rather than ontology traversal. The paper explicitly describes a dynamically expanded query set built from meme text, image description, and generated concepts:

$Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$

The printed LaTeX is malformed in the paper, but the intended symbols are clear in the accompanying explanation: $T$ denotes meme text, $X$ the image description, $T_{\text{set}}$ the keyword or concept set generated by a lightweight LLM, and $Q_{\text{set}}$ the combined retrieval query set.

Retrieval then uses a hybrid sparse+dense score:

$S_{\text{hybrid}(d,q) = w_{\text{bm25} \cdot S_{\text{BM25}(d,q) + w_{\text{dense} \cdot S_{\text{dense}(d,q)$

Again, the typesetting is broken, but the intended semantics are explicit: $S_{\text{BM25}}$ provides lexical matching, $S_{\text{dense}}$ provides dense semantic similarity, and $w_{\text{bm25}}, w_{\text{dense}}$ balance them. This is important because Chinese harmful expressions often depend on both exact surface forms and broader semantic similarity.

The reranker is described as computing a binary yes/no relevance posterior:

$Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$0

with thresholding:

$Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$1

Only documents above the threshold are retained. A common misconception is that C-HarmKB is introduced as a knowledge graph of triples. The paper does not do this. It introduces a textual knowledge repository with vectorized entries, hybrid retrieval, and reranking (Wang et al., 23 May 2026).

4. Function inside RIKE and Ex-ToxiCN-MM

C-HarmKB is operationalized through the RIKE framework, which has two stages: Attribution Knowledge Enhancement (AKE) and Relative Intent Reasoning (RIR). The meme is represented as

$Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$2

where $Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$3 is the image, $Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$4 is embedded or OCR text, and $Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$5 is a textual image description. Given $Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$6, AKE retrieves the top-$Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$7 relevant knowledge fragments, denoted $Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$8, from C-HarmKB (Wang et al., 23 May 2026).

The role of Ex-ToxiCN-MM is to provide paired interpretations for the same meme:

• $Q_{\text{set} = \{q|q \in (T \cup X \cup \{T_{\text{set}\}) \cap q \neq \emptyset\}$9: non-harmful interpretation
• $T$0: harmful interpretation

RIR then reasons over the meme, the two opposing interpretations, and the retrieved knowledge. The final decision is defined as:

$T$1

The classification objective is given as

$T$2

and the explanation-generation objective as

$T$3

with the caveat that the printed equations are slightly malformed in the paper. The substantive point is unambiguous: retrieved background knowledge conditions both classification and explanation generation.

At the prompt level, RIR explicitly includes:

• Meme Image: $T$4
• Meme Text: $T$5
• Meme Image Description: $T$6
• Meme Background Knowledge: $T$7
• Interpretation of Non-harmful Perspective: $T$8
• Interpretation of Harmful Perspective: $T$9

This architecture makes C-HarmKB not merely a retrieval aid but explicit evidence in attribution reasoning. The paper’s own summary is that AKE solves knowledge scarcity and RIR solves attribution ambiguity. In that formulation, C-HarmKB is the knowledge backbone of interpretable Chinese harmful meme attribution (Wang et al., 23 May 2026).

5. Relation to adjacent Chinese harmful-content resources

C-HarmKB belongs to a broader ecosystem of Chinese harmful-content resources, but occupies a specific niche. ToxiCN introduced Monitor Toxic Frame, a hierarchical taxonomy separating general offensive language from hate speech, with target-group labels and expression categories (explicitness, implicitness, reporting), alongside a 1,032-term insult lexicon and Toxic Knowledge Enhancement (TKE) (Lu et al., 2023). That resource is text-centric and taxonomy-driven; C-HarmKB is meme-centric and oriented toward cultural disambiguation in multimodal attribution.

STATE ToxiCN shifted Chinese hate-speech analysis to span-level structure with Target-Argument-Hateful-Group quadruples over 8,029 posts and 9,533 quadruples, and also provided an annotated hateful slang resource. One account describes a 330-term hateful slang lexicon (Bai et al., 26 Jan 2025); a later account describes an 830-term Chinese coded hate term lexicon with JSON schema $X$0 and categories such as Phono-Graphemic Variation Term and Hateful Semantic Term (Bai et al., 15 Jul 2025). These resources are especially relevant because they expose coded, homophonic, and historically allusive Chinese hate expressions that a broader C-HarmKB could plausibly absorb.

ChineseHarm-Bench broadens the problem space beyond toxicity and hate speech to Gambling, Pornography, Abuse, Fraud, Illicit Ads, and Non-violation, and its annotation process yields a knowledge rule base $X$1 of category-specific moderation rules (Liu et al., 12 Jun 2025). Relative to that benchmark, C-HarmKB is narrower in modality and task scope, but deeper in meme-specific cultural semantics.

Two later directions show how KB-style knowledge is being operationalized. Chinese Implicit Toxicity Attack (CITA) distinguishes harmful intent, implicit toxicity enhancement, and obfuscation rewriting, with four obfuscation types: Homophone Replacement, Character Transposition, Traditional Mixing, and Emoji-based Substitution (Kang et al., 21 May 2026). KID formulates harmful meme detection as entity-anchored knowledge construction and dual-head learning, using interleaved descriptions of the form $X$2 (Li et al., 29 Jan 2026). Taken together, these works suggest that Chinese harmful-content research is moving from label-only prediction toward explicit semantic resources, rule bases, lexicons, and grounded knowledge injection.

6. Empirical significance, misconceptions, and open issues

The empirical case for C-HarmKB is given indirectly through background-knowledge augmentation in RIKE. In the harmful meme attribution task, adding background knowledge improves performance consistently. For Qwen2.5-VL-3B, RIR achieves Acc. 0.566, P 0.567, R 0.772, F1 0.654, while RIR+BK reaches Acc. 0.619, P 0.595, R 0.878, F1 0.710. For Qwen2.5-VL-7B, the corresponding change is 0.643/0.805/0.431/0.562 to 0.702/0.834/0.546/0.660. For InternVL3_5-8B, it is 0.614/0.810/0.355/0.494 to 0.674/0.840/0.477/0.608. For LLaVA-1.5-7b, it is 0.557/0.583/0.580/0.581 to 0.592/0.607/0.655/0.630 (Wang et al., 23 May 2026).

Qualitative cases are equally central. The paper states that retrieved contextual knowledge corrected baseline deficiencies in understanding cultural context: “working like a horse” was interpreted as labor intensity rather than a generic life statement; “Asian parents” was recognized as a stereotype-bearing concept; and a violent scene was no longer hallucinated as “mouth-to-mouth resuscitation.” These examples support the claim that C-HarmKB reduces false negatives and shallow explanations by grounding models in Chinese cultural semantics (Wang et al., 23 May 2026).

Several misconceptions are addressed by the paper itself. C-HarmKB is not presented as a symbolic graph with explicit relation types. It is not only an offensive-vocabulary list. It is not a sufficient solution by itself. Remaining failure cases show that even with retrieved knowledge, models may overweight the harmless perspective, overlook psychological harm, or treat despair-laden self-deprecating humor as harmless. This suggests that knowledge retrieval, attribution weighting, and normative judgment remain distinct problems.

The major open issues are also explicit. The paper does not provide a formal schema, per-category counts, relation types, inter-annotator agreement for KB annotation, source-by-source statistics, or coverage rates over Ex-ToxiCN-MM (Wang et al., 23 May 2026). A plausible implication is that future work will need to formalize C-HarmKB beyond textual entries in a vector database. The surrounding literature already points toward likely directions: span-level target-aware harmful semantics and coded-term lexicons (Bai et al., 15 Jul 2025), broader rule-based harmful-content coverage (Liu et al., 12 Jun 2025), multimodal rationale resources with chain-of-thought annotations (Gu et al., 15 Jun 2025), and explicit modeling of implicit enhancement and obfuscation (Kang et al., 21 May 2026). Within that trajectory, C-HarmKB is best understood as the first published attempt to make Chinese harmful meme attribution depend on an explicit, reusable store of cultural and offensive semantics rather than on latent model knowledge alone.