Conflict Attribution Corpus (CAC)

• CAC is defined as two distinct resources: a Reddit-based corpus for studying interpersonal conflict and a multimodal dataset for explicit contradiction attribution.
• The AITA-based CAC uses six annotated conflict dimensions and SBERT clustering to evaluate social judgments (YTA vs. NTA) across diverse relationship types.
• The multimodal CAC in the CORE framework employs structured conflict supervision with cross-modal alignment to enhance reasoning against manipulated information.

Searching arXiv for the specified papers and topic to ground the article in the cited literature. arxiv_search({"query":"\"Conflict Attribution Corpus\" OR (Welch et al., 2022) OR (Shen et al., 2 Jun 2026)","max_results":10,"sort_by":"relevance"}) The Conflict Attribution Corpus (CAC) denotes two distinct research corpora that use the same acronym in different domains. In computational social science, CAC is a Reddit-based dataset centered on the AITA (“Am I the Asshole”) subreddit and was introduced to study interpersonal conflict, conflict typology, and perception classification of YTA versus NTA verdicts (Welch et al., 2022). In multimodal misinformation detection, CAC is the central supervision resource of the CORE framework and is defined as a corpus of image-text pairs annotated with explicit contradictory content and its sources, so that multimodal LLMs can learn conflict-oriented reasoning rather than manipulation-specific heuristics (Shen et al., 2 Jun 2026). The shared term “conflict attribution” therefore refers not to a single canonical dataset, but to two independently motivated annotation schemes that make latent conflict structure explicit for downstream modeling.

1. Interpersonal-conflict CAC in AITA research

The interpersonal-conflict CAC was introduced as a new Reddit-based resource for studying how people perceive and judge interpersonal conflict, and it is built around the AITA subreddit, where an original poster describes a conflict and community members judge whether the author’s behavior was wrong or not (Welch et al., 2022). Its stated contribution is twofold: it adds a conflict-typology annotation layer grounded in interpersonal conflict literature, and it supports experiments on perception classification, defined as predicting whether a verdict writer will judge the author as YTA (“you’re the asshole”) or NTA (“not the asshole”) (Welch et al., 2022).

The corpus was created because conflict judgments are deeply tied to social norms, but those norms are not uniform, and existing computational work on AITA mainly modeled verdict prediction without explicitly connecting posts to conflict theory (Welch et al., 2022). CAC was designed to bridge that gap by providing a large AITA corpus for modeling verdicts, a manually annotated subset with conflict-aspect labels derived from conflict psychology and organizational conflict literature, automatically induced clusters of conflict situations that were human-validated as meaningful groupings, and analysis of how conflict type and social relation affect model performance and human judgments (Welch et al., 2022).

The raw dataset consists of 21K posts and 364K verdict comments, including 254K NTA and 110K YTA, and all data are in English (Welch et al., 2022). For modeling, the resource uses the post title (“situation”) and/or the full post text, and the released code and dataset include the 21K posts, 364K comments, two sets of cluster labels, and 500 posts annotated with six conflict aspects, corresponding to 1,653 verdicts (Welch et al., 2022).

This design makes CAC more than a verdict-prediction benchmark. It is a corpus intended to connect computational modeling of moral judgment to structured conflict theory, with the explicit goal of explaining not just whether a comment is judged YTA or NTA, but why certain conflict scenarios are easier or harder to judge (Welch et al., 2022).

2. Annotation scheme and theoretical grounding

The annotation scheme of the AITA-based CAC is explicitly derived from prior conflict theory. The paper cites Barki & Hartwick (2004), who characterize interpersonal conflict in terms of disagreement, negative emotion, and interference; Bendersky et al. (2014), who add duration and manifestation vs. perception; and Korsgaard et al. (2008), who motivate number of people involved / group-level conflict (Welch et al., 2022). These were refined into a six-aspect conflict annotation scheme:

1. Strength of disagreement
2. Intensity of negative emotion
3. Degree of interference
4. Duration of conflict
5. Manifestation vs. perception
6. Number of people involved (Welch et al., 2022)

Some of the annotation questions originally had 3-way scales, but for analysis the authors merged labels into binary categories when needed (Welch et al., 2022). For disagreement strength and negative emotion intensity, Strong + Intense → Strong. For degree of interference, Not at all + Somewhat → Mild, while Strongly → Strong. Duration is annotated as One-time incident or Longer. Manifestation vs. perception distinguishes Manifest conflicts, expressed in observable actions or speech, from Perceived conflicts that exist mainly in someone’s mind. Number of people involved is labeled One person or Multiple people (Welch et al., 2022).

The annotation procedure began with 25 posts used to refine the scheme because existing conflict scales were found not well suited to AITA-style conflicts: not all conflicts are workplace conflicts, distinctions such as “friction vs. tension vs. emotional conflict” were too fine-grained for the setting, and long-term conflict is often hard to infer from short narratives (Welch et al., 2022). The larger annotation phase covered 500 posts and 1,653 comments/verdicts, using 14 annotators total, recruited via Prolific and university researchers; all had English fluency and all passed two attention checks in every survey (Welch et al., 2022).

Inter-annotator agreement was measured with Matthews correlation coefficient (MCC). Reported MCCs were 0.39 → 0.49 after label merging for disagreement strength, 0.33 → 0.41 for emotion intensity, 0.13 → 0.20 for interference degree, 0.39 for conflict duration, 0.10 for manifestation or perception, and 0.40 for number of people (Welch et al., 2022). The strongest agreement was for disagreement, duration, and number of people, while the weakest was for manifestation vs. perception and interference (Welch et al., 2022). This indicates that some conflict dimensions are operationally easier to identify in short social narratives than others.

The merged label distribution was reported as 33.0% Mild, 67.0% Strong for disagreement; 35.7% Mild, 64.3% Strong for emotion; 35.3% Mild, 64.7% Strong for interference; 48.3% Once, 51.7% Longer for duration; 33.7% Perceived, 66.3% Manifest for manifestation; and 72.0% One, 28.0% More for number of people (Welch et al., 2022). A plausible implication is that the corpus is skewed toward relatively strong, explicit, and dyadic conflicts, which is consistent with the narrative style of AITA posts, but that interpretation remains secondary to the reported distributions themselves.

3. Clustering and relationship-structured conflict types

Before manual annotation, the AITA-based CAC explored whether AITA conflicts naturally group into types (Welch et al., 2022). The authors used two text representations: situations, defined as the post title, and full text, defined as the entire post body (Welch et al., 2022). They computed Sentence-BERT (SBERT) embeddings, pairwise cosine similarity normalized to $[0,1]$, formed a fully connected weighted graph, and applied Louvain clustering to maximize modularity (Welch et al., 2022).

Weak edges were pruned by dropping the bottom N% of edge weights, where $N$ was chosen based on stability measured by adjusted Rand index (ARI) across nearby thresholds (Welch et al., 2022). The chosen pruning levels were 40% for situations and 30% for full text. This yielded 3 clusters for situations and 3 clusters for full text, after removing one tiny 4th cluster from the 30% full-text cutoff because it had only 25 posts (Welch et al., 2022).

Manual inspection and human validation showed that the clusters mainly reflect the social relation of the author to the other participant(s): Family, Close relationships, and Distant relationships (Welch et al., 2022). The authors emphasize that the subject of the conflict is strongly entangled with relationship closeness, and that relation boundaries themselves can be ambiguous (Welch et al., 2022). In human validation, two authors manually clustered 100 posts and found disagreement sources in how to draw boundaries between close relationship types and in ambiguity about who is actually involved in the conflict (Welch et al., 2022). Reported ARI values were 0.33 between the two humans, 0.38 and 0.15 between full-text clusters and humans, and 0.31 and 0.13 between situation clusters and humans (Welch et al., 2022). The clustering is therefore only moderately aligned with manual labels, but the categories were still described as meaningful and analytically useful (Welch et al., 2022).

This relationship-structured interpretation is central to the corpus’s analytical use. The paper’s main takeaway is that conflict content varies systematically with social relationship, and that this affects both how humans judge a situation and how well a model can predict those judgments (Welch et al., 2022). The reported gradient is clear: family conflicts are easiest to model, close relationships are intermediate, and distant relationships are hardest (Welch et al., 2022). The authors interpret this as an indirect relationship between relational closeness and prediction difficulty, with closer relationships tending to produce richer, more explicit conflict narratives that are easier for the model to classify (Welch et al., 2022).

4. Perception classification and empirical findings

The core modeling task in the AITA-based CAC is to predict whether an observer judges the author’s action as right or wrong, operationalized as NTA vs. YTA (Welch et al., 2022). The model operates on the comment text together with the situation title by concatenating situation + comment, after removing explicit verdict labels from the comment text (Welch et al., 2022). The main model is a fine-tuned SBERT classifier, and the authors also report that adding a representation of the full text as extra features did not improve performance over using the comment plus situation alone (Welch et al., 2022). The baseline is JudgeBERT from Botzer et al. (2022), a BERT-base classifier with dropout and a classification layer, re-implemented on this dataset (Welch et al., 2022). Training used 10 epochs, Adam, learning rate 1e-4, and focal loss to cope with imbalance between YTA and NTA (Welch et al., 2022).

The data for perception classification were split 70/20/10 into train/validation/test, stratified by situation clusters and by full-text clusters (Welch et al., 2022). The annotated subset consists of 500 posts from the test set, with 1,653 comments/verdicts annotated for conflict aspects (Welch et al., 2022).

The SBERT-based approach outperformed JudgeBERT on both stratifications. Under full text stratification, JudgeBERT achieved 72.7 F1, 84.9 Acc, while the proposed model achieved 77.2 F1, 87.0 Acc. Under situation stratification, JudgeBERT achieved 70.1 F1, 83.2 Acc, while the proposed model achieved 77.4 F1, 87.2 Acc (Welch et al., 2022). The paper reports that the improvement is statistically significant by permutation test, p < 0.0001, with about +5 F1 on full text and about +7 F1 on situations (Welch et al., 2022).

Using full-text clusters, the model also outperformed JudgeBERT in all three relationship clusters: Family (74.9 F1 / 86.8 Acc vs. 79.0 F1 / 88.3 Acc), Close (72.2 F1 / 84.4 Acc vs. 76.7 F1 / 86.9 Acc), and Distant (71.2 F1 / 82.2 Acc vs. 75.9 F1 / 85.0 Acc) (Welch et al., 2022). The situation-stratified numbers show the same pattern, with best performance on Family, lower on Close, and lowest on Distant (Welch et al., 2022).

The paper’s most distinctive analysis compares model performance across the six conflict dimensions. The reported trends are that more negative emotion, stronger disagreement, and more people involved make classification harder, while stronger interference, longer duration, and more manifest conflicts make classification easier (Welch et al., 2022). Under full-text stratification, the reported aspect-level results were:

Aspect	Condition	Reported results
Disagreement	Mild	89.5 Acc / 70.8 micro F1 / 78.0 macro F1
Disagreement	Strong	88.3 Acc / 69.5 micro F1 / 76.4 macro F1
Emotion	Mild	88.3 / 70.0 / 77.8
Emotion	Strong	84.0 / 69.6 / 76.6
Interference	Weak	84.7 / 56.4 / 74.5
Interference	Strong	86.3 / 85.5 / 85.5
Duration	Once	82.7 / 68.2 / 71.7
Duration	Longer	86.5 / 70.7 / 82.0
Manifestation	Perceived	81.8 / 51.9 / 73.2
Manifestation	Manifest	86.4 / 73.7 / 78.9
Number of people	One	86.1 / 73.1 / 78.5
Number of people	More	80.0 / 42.5 / 72.7

Most dyads differed significantly under one-sided unpaired permutation tests, except interference, where improvement was not statistically significant in the model comparison (Welch et al., 2022). In the appendix, the authors also analyze how verdict distributions vary across conflict dimensions using Fisher’s exact test. They report that Strong disagreement has an 11% higher YTA/NTA ratio difference than mild, Strong emotion differs by 9%, Strong interference shows the largest difference at 78%, One-time incidents have a 13% higher YTA/NTA ratio difference, More manifest conflicts show a 13% higher difference, and Single-person conflicts differ by 11% (Welch et al., 2022).

These results support the paper’s broader claim that “social norm violations” are not a single uniform phenomenon, but depend on emotional intensity, disagreement strength, interference, duration, explicitness, and relationship structure (Welch et al., 2022). This suggests that the AITA-based CAC is both a dataset and an empirical argument for modeling moral judgment as conflict-sensitive rather than label-only.

5. Multimodal-conflict CAC in CORE

A separate corpus with the same acronym appears in the CORE framework for multimodal fake news and manipulation detection (Shen et al., 2 Jun 2026). In that setting, the Conflict Attribution Corpus (CAC) is the paper’s central supervision resource for teaching multimodal LLMs to perform conflict-oriented reasoning rather than relying on manipulation-specific cues (Shen et al., 2 Jun 2026). The stated motivation is that existing detection methods rely heavily on manipulation-specific models and large-scale labeled data, which leads to poor generalization to emerging manipulation types (Shen et al., 2 Jun 2026). The authors’ core observation is that manipulated multimodal misinformation is unified by an underlying conflict—a semantic contradiction or a physical inconsistency appearing either across modalities or between a sample and world knowledge (Shen et al., 2 Jun 2026).

CAC in this setting was created because existing benchmarks and training sets do not provide the fine-grained conflict supervision needed for a model to learn human-like deception detection (Shen et al., 2 Jun 2026). Rather than supplying only real/fake labels or manipulation categories, it annotates manipulated samples with the exact contradictory content and its origin, enabling a model to learn the abstract notion of conflict itself (Shen et al., 2 Jun 2026). The corpus is therefore explicitly designed for robustness to out-of-distribution and unseen manipulation types (Shen et al., 2 Jun 2026).

The corpus is built from 100k image-text pairs selected from SAMM as the source pool (Shen et al., 2 Jun 2026). The construction pipeline has four steps: Source Sample Selection, Background Knowledge Collection, Conflict Rationale Generation, and Conflict Structuring (Shen et al., 2 Jun 2026). Background knowledge is retrieved using the Google Search API. The image, caption, manipulation prior, and background knowledge are then fed into an MLLM randomly chosen from $$\{GPT-4o,\ \text{Gemini2.5-Pro},\ \text{Qwen3-VL-Plus}\}$$, and the generated rationale is cross-validated by the other two MLLMs (Shen et al., 2 Jun 2026). The validated rationale is distilled into a structured annotation with two Conflict Factors and their corresponding Conflict Sources (Shen et al., 2 Jun 2026).

The paper defines Conflict Factor (C) as the specific contradictory content, with examples such as “US President”, “Ballon d’Or”, “normal face”, and “unnatural face skin”, and defines Conflict Source (S) as where that contradictory element comes from, with allowed source labels Image, Caption, and World Knowledge (Shen et al., 2 Jun 2026). The final record format is

$$\langle \text{Image } I, \text{Text } T, \{C_1, C_2\}, \{S_1, S_2\} \rangle.$$

This schema makes the multimodal CAC a dataset of explicit contradiction attribution rather than a binary label set (Shen et al., 2 Jun 2026).

After validation, the paper reports that CAC contains 14k instances (Shen et al., 2 Jun 2026). The source distribution is 29.98% from the caption, 36.86% from the image, and 33.16% from world knowledge (Shen et al., 2 Jun 2026). Human verification was performed by 5 annotators on a random sample of 1k items, and 993 samples passed human review, yielding a 99.3% pass rate (Shen et al., 2 Jun 2026). The labeling rules emphasize Conflict existence, Source accuracy, and Granularity, requiring a real logical contradiction, correct attribution to Image, Caption, or World Knowledge, and fine-grained concepts rather than vague labels (Shen et al., 2 Jun 2026).

6. Conflict types, supervision role, and relation between the two CACs

The multimodal CAC captures multiple conflict types: semantic conflicts, physical inconsistencies, cross-modal conflicts, and world-knowledge conflicts (Shen et al., 2 Jun 2026). Examples include “Donald Trump won the football award,” where the contradiction is between the caption and world knowledge; mismatches between a caption describing a politician and an image depicting an unrelated celebrity or athlete; and visual inconsistencies such as lighting or shadows, facial skin artifacts, unnatural symmetry, blurred or sharpened regions, and mismatched emotional expressions (Shen et al., 2 Jun 2026). The paper’s examples include annotations such as CF: actress - Everton attack, CS: WK - Caption, CF: basketball player - Perform, CS: WK - Image, CF: TV Show - football club, CS: Image - Caption, and CF: football coach - government official, CS: WK - Caption (Shen et al., 2 Jun 2026).

CAC is the supervision backbone of CORE’s Conflict Perception Training (CPT) (Shen et al., 2 Jun 2026). Because CAC annotations are textual but some conflict sources originate from images, CORE first learns a Cross-modal Aligner during Modality Bridging Pre-Training (MBPT). The aligner is trained with FineHARD using

$$V_p = \text{Aligner}(t_p, V, V), \quad \{v_{n_i}\} = \text{Aligner}(\{t_{n_i}\}, V, V) \tag{1}$$

and

$$L_a = \mathbb{E}_{(t,v)\in Q} \frac{1}{1 + e^{y_t(s_1\cdot (t,v) + b_1)}} \tag{2}$$

with total MBPT loss

$L_{\text{mbpt}} = L_{cl} + L_{o2vqa}. \tag{3}$

In Conflict Perception Training, given a CAC sample

$\langle I, T, \{C_1, C_2\}, \{S_1, S_2\} \rangle,$

the two conflict factors are converted into features $t_{c1}, t_{c2}$, and if a factor comes from the image, the aligned visual feature is used: $$z_i = \begin{cases} \text{Aligner}(t_{c_i}, V, V), & \text{if } S_i \text{ is image} \ t_{c_i}, & \text{otherwise} \end{cases} \tag{4}$$ CORE then applies a conflict-aware contrastive objective

$N$0

and adds a conflict reasoning objective with the target verbalization:

“Real. / Fake. Because the $N$1 from $N$2 conflicts with $N$3 from $N$4.” The total CPT loss is

$N$5

All of these components are described as mechanisms for making MLLMs not just know facts, but learn to separate conflicting concepts in feature space and verbalize the contradiction (Shen et al., 2 Jun 2026).

Although the interpersonal-conflict CAC and the multimodal-conflict CAC are unrelated datasets in provenance, they share a methodological theme: both attempt to convert a latent, human-interpretable notion of conflict into explicit annotation structure. In the AITA-based CAC, the target is conflict typology and perceived right/wrong judgment in interpersonal situations (Welch et al., 2022). In the CORE CAC, the target is contradiction attribution in manipulated multimodal information (Shen et al., 2 Jun 2026). This suggests a common editorial shorthand—“structured conflict supervision” (Editor's term)—for their shared design principle, though the two corpora differ sharply in modality, ontology, and downstream task.

7. Scope, distinctions, and common misconceptions

A common misconception is that Conflict Attribution Corpus (CAC) refers to a single established benchmark. The available arXiv literature instead documents at least two separate resources with that name: one for interpersonal conflict and social-norm judgment in AITA (Welch et al., 2022), and one for multimodal contradiction attribution in fake-news detection (Shen et al., 2 Jun 2026). Their overlap is nominal rather than genealogical.

Another possible confusion is with other uses of the acronym CAC in unrelated literatures. In reverse mathematics, CAC denotes the ordinary chain-antichain theorem, and “CAC for trees” is the statement that every infinite subtree of $N$6 has an infinite path or an infinite antichain (Cervelle et al., 2022). In coding theory, CAC denotes conflict-avoiding codes, a deterministic transmission scheme for asynchronous multiple access without feedback (Lo et al., 2020, Lo et al., 2024). These meanings are unrelated to either corpus. The cyber-attribution paper “Argumentation Models for Cyber Attribution” uses a DEFCON CTF attribution dataset, but it does not introduce a dataset called Conflict Attribution Corpus by that name (Nunes et al., 2016).

Within the social-computing CAC, an additional misconception would be to treat the six conflict aspects as if they were uniformly easy to annotate. The reported MCC values show otherwise, especially for manifestation vs. perception and interference, which were among the hardest aspects to annotate (Welch et al., 2022). Within the multimodal CAC, a parallel misconception would be to regard it as merely a real/fake benchmark. Its defining contribution is the structured annotation of Conflict Factors and Conflict Sources, including World Knowledge as an explicit source category (Shen et al., 2 Jun 2026).

Taken together, the corpora show two different ways of operationalizing conflict for machine learning. One treats conflict as a socially embedded situation whose dimensions affect human judgment and prediction difficulty (Welch et al., 2022). The other treats conflict as explicit contradiction structure that can supervise reasoning under unseen manipulations (Shen et al., 2 Jun 2026). A plausible implication is that “conflict attribution” has emerged as a cross-domain design pattern for moving beyond coarse labels toward explanatory supervision, but the specific ontologies remain domain-dependent.