GET /cau/{id}: 특정 주장에 대한 신뢰도 평가 결과와 근거 자료 상세 조회GET /causal/graphs: 특정 시간대별 여론 확산의 인과 그래프 시각화 데이터 조회POST /interventions/simulate: 특정 조치(라벨링 등) 적용 시 예상 파급 효과 시뮬레이션 요청Dependency parsing + constituency parsing으로 절 경계, 주절/종속절, 인용절 구조를 복원
negation, modality, attribution, quantifier의 적용 범위를 계산해 proposition에 부착
평가 가능한 절/내포절만 분리해 claim candidate 또는 evidence candidate로 전환
Sentence Structure Parsing
constituency / dependency / clause segmentation / scope resolution
Semantic Chunk & Proposition Layer
surface chunk + proposition unit + operator tagging
Epistemic Composition Layer
belief, uncertainty, attribution-weight 계산 후 CAU 생성
| 연산자 유형 | 예시 | 언어학적 역할 | 시스템 처리 |
|---|---|---|---|
| Attribution | “전문가들은 …라고 말했다” | 발화 주체와 인용 내용을 분리 | claim 직접성 감소, source credibility와 분리 계산 |
| Modal / Hedge | may, might, ~일 수 있다, ~로 보인다 | 확실성 하향, 추정/가설 표시 | belief weight downscale, uncertainty 증가 |
| Negation | not, no evidence, 아니다, 못하다 | 명제 방향성 및 스코프 반전 | scope 기반 polarity/stance 재계산 |
| Evidentiality | ~라고 한다, ~로 알려졌다 | 직접 관찰이 아닌 전언/보도 표지 | reported speech로 분리, evidence prior 조정 |
| Condition / Concession | 만약 ~라면, ~에도 불구하고 | 가정/양보 구조로 direct fact claim과 구분 | hypothetical claim 또는 context proposition으로 분류 |
claim별 belief, uncertainty, stance를 산출
시점별 epistemic 상태를 요약한 feature vector 생성
확산·영향 DAG의 입력 변수로 사용
직접 엣지를 연결하지 말고, 아래와 같은 중간 객체를 두는 것이 DB/시스템 설계상 가장 안전합니다.
이 변수들은 claim 자체가 아니라, 현재 그 claim이 어떤 epistemic 상태에 놓여 있는지를 요약합니다.
Most claim verification systems treat claims as flat sentences and ignore the internal sentence structure that shapes epistemic meaning. In this work, we propose a linguistically grounded framework that models claims as structured propositions derived from sentence structure, clause hierarchy, and epistemic operators such as attribution, modality, negation, and evidentiality. Our approach decomposes sentences into proposition units and composes them into Claim Assessment Units (CAUs), enabling explicit modeling of belief, uncertainty, and rationale. Experiments on misinformation and fact‑verification datasets show that incorporating sentence structure and epistemic operators improves evidence reasoning and claim verification performance compared to sentence‑level baselines. We also introduce an annotation schema and dataset design tailored for epistemic claim modeling in misinformation analysis.
| 구분 | 내용 |
|---|---|
| Baseline | sentence-level encoder, retrieval+NLI, operator 미반영 모델 |
| Proposed | structure-aware proposition encoder + epistemic operator composition |
| Ablation | sentence structure 제거 / operator 제거 / source prior 제거 / uncertainty head 제거 |
| Tasks | evidence retrieval, stance classification, claim verification, uncertainty calibration |
| 평가 항목 | 지표 |
|---|---|
| Evidence Retrieval | Recall@k, Precision@k, MRR |
| Stance Classification | Macro-F1, Accuracy |
| Claim Verification | Label Accuracy, Macro-F1 |
| Uncertainty | ECE, Brier Score, calibration plots |
Most claim verification systems treat claims as flat sentences and ignore the internal sentence structure that shapes epistemic meaning. In this work, we propose a linguistically grounded framework that models claims as structured propositions derived from sentence structure, clause hierarchy, and epistemic operators such as attribution, modality, negation, and evidentiality. Our approach decomposes sentences into proposition units and composes them into Claim Assessment Units (CAUs), enabling explicit modeling of belief, uncertainty, and rationale. Experiments on misinformation and fact‑verification datasets show that incorporating sentence structure and epistemic operators improves evidence reasoning and claim verification performance compared to sentence‑level baselines. We also introduce an annotation schema and dataset design tailored for epistemic claim modeling in misinformation analysis.
Fact verification tasks such as FEVER typically model claim–evidence reasoning using sentence‑pair classification pipelines. However, real-world misinformation often involves multiple pieces of evidence with varying credibility and contradictory signals. We propose an epistemic graph framework that represents claim–evidence relationships as structured graphs and aggregates them into an epistemic state capturing belief, uncertainty, contradiction ratio, and evidence density. This representation enables graph‑based evidence reasoning and improves interpretability by producing rationale subgraphs that explain verification decisions. Experiments demonstrate that incorporating epistemic state features enhances evidence selection, stance prediction, and verification robustness compared to standard FEVER pipelines. We further present a dataset design compatible with FEVER-style evaluation while supporting richer epistemic annotations.
| 구분 | 내용 |
|---|---|
| Baseline | 표준 FEVER pipeline, sentence pair NLI, reranker 기반 evidence selection |
| Proposed | graph-based evidence reasoning + epistemic state aggregation |
| Ablation | graph 제거 / epistemic state 제거 / source credibility 제거 / rationale subgraph 제거 |
| Tasks | evidence selection, multi-evidence reasoning, label prediction, rationale explanation |
| 평가 항목 | 지표 |
|---|---|
| Evidence Selection | FEVER evidence score, Recall@k |
| Label Prediction | FEVER score, Accuracy, Macro-F1 |
| Graph Quality | edge-level F1, rationale overlap |
| Explainability | human preference, rationale faithfulness |
Recent advances in natural language processing have significantly improved automated fact verification and misinformation detection. Most existing approaches treat claim verification as a sentence‑pair classification problem, where a claim and a candidate evidence sentence are encoded and evaluated using neural entailment models. While effective in controlled benchmarks, such approaches assume that claims can be represented as flat sentences and that evidence reasoning occurs independently across sentence pairs. However, real‑world claims often contain complex linguistic structures—including attribution (“experts say”), modality (“may cause”), negation (“not proven”), and evidential markers (“according to reports”)—that fundamentally alter the epistemic meaning of the statement. Ignoring these structures can obscure what is actually being asserted, who is making the assertion, and how strongly the claim is expressed.
In this work, we argue that claim verification should be grounded in sentence structure and epistemic interpretation. We propose a linguistically informed framework that decomposes sentences into clause‑level proposition units and explicitly models epistemic operators such as attribution, modality, negation, and evidentiality. These elements are combined into structured Claim Assessment Units (CAUs) that capture not only the semantic content of a claim but also its epistemic context. By representing claims and evidence within a structured claim–evidence graph, our framework enables richer reasoning over multiple pieces of evidence and supports the computation of epistemic state variables such as belief strength, uncertainty, contradiction ratio, and evidence density.
We evaluate the proposed approach on misinformation and fact‑verification datasets and compare it against strong sentence‑level baselines. Our experiments show that incorporating sentence structure and epistemic operators improves evidence reasoning and claim verification performance, while also enabling calibrated uncertainty estimates and interpretable rationale subgraphs. Furthermore, we introduce an annotation schema and dataset design tailored for epistemic claim modeling, which facilitates future research on linguistically grounded verification. Overall, this work demonstrates that integrating linguistic structure with epistemic reasoning provides a promising direction for more robust and interpretable fact verification systems.
Clause Graph는 문장을 clause-level proposition의 그래프로 표현한 구조입니다. 각 node는 절(clause)을 나타내고, edge는 attribution, condition, contrast, complement 등의 관계를 나타냅니다.
여기서 Clause 2가 claim candidate이고, Clause 1은 epistemic modifier 역할을 합니다.
“Experts say ...”
발화 주체와 주장 내용을 분리하며, direct assertion보다 claim strength를 낮춥니다.
“may”, “might”, “~일 수 있다”
가능성/추정 표지로 belief를 낮추고 uncertainty를 높입니다.
“not proven”, “according to reports”
부정 스코프와 정보 출처를 반영하여 proposition의 해석을 조정합니다.
Complex claim sentence with attribution, modality, negation, and evidential markers
Dependency / constituency parsing, clause segmentation, scope resolution
Clause-level nodes linked by attribution, condition, complement, or contrast
Extract verifiable proposition units from clause graph
Attach attribution, modality, negation, evidentiality to propositions
Structured claim object for belief, uncertainty, and rationale computation
Retrieve candidate evidence and link them to CAUs with stance edges
Graph over claims and evidence with support / contradict / uncertain relations
Aggregate belief, uncertainty, contradiction ratio, evidence density, credibility
SUPPORTED / REFUTED / NEI + calibrated uncertainty + rationale subgraph
| 시스템 모듈 | 대응 연구개념 | 주요 역할 | 입력 / 출력 | 우선순위 | 비고 |
|---|---|---|---|---|---|
| sentence_structure_parser | Sentence Structure Parsing Layer | 문장 분리, dependency/constituency parsing, 절 경계 탐지, scope 해석 후보 생성 | 입력: 원문 텍스트 출력: sentence_tree, clause_graph, operator_scope_map |
즉시 | 전처리 핵심 엔진 |
| clause_graph_builder | Clause Graph | 주절/종속절/인용절 관계를 그래프로 구성하고 attribution·condition·contrast 등 edge 부여 | 입력: parse 결과 출력: clause_graph |
즉시 | ACL 핵심 novelty |
| proposition_extractor | Proposition Units | 검증 가능한 절 수준 명제 추출, claim candidate / context candidate 구분 | 입력: clause_graph 출력: proposition_units, normalized_proposition |
즉시 | rule + parsing 병행 가능 |
| epistemic_operator_tagger | Epistemic Operators | attribution, modality, negation, evidentiality, condition/concession 탐지 및 스코프 부착 | 입력: proposition, scope map 출력: operator_bundle, epistemic modifiers |
즉시 | 시연 효과 큼 |
| cau_builder | Claim Assessment Unit (CAU) | proposition + operator + source context를 통합하여 claim 평가 기본 객체 생성 | 입력: proposition, operators, metadata 출력: cau_id, claim_repr, belief_prior, uncertainty_prior |
즉시 | DB 중심 핵심 객체 |
| evidence_retriever | Evidence Retrieval | claim 기준 외부 근거 검색, relevance ranking, 출처 메타 수집 | 입력: CAU 출력: evidence candidates, retrieval score |
핵심 | RAG/검색 결합 |
| stance_classifier | Support / Contradict Reasoning | evidence가 claim을 지지/반박/불확실 중 무엇인지 판정 | 입력: CAU + evidence 출력: stance label, confidence |
핵심 | FEVER 직접 연결 |
| claim_evidence_graph_builder | Claim–Evidence Graph | claim/evidence 노드와 support·contradict·uncertain edge를 구성 | 입력: CAU, evidence, stance 출력: graph json / graph tables |
핵심 | 시각화 친화적 |
| epistemic_state_aggregator | Epistemic State | belief, uncertainty, contradiction ratio, evidence density, credibility를 요약 계산 | 입력: claim–evidence graph 출력: epistemic snapshot |
핵심 | 대시보드 지표화 가능 |
| verification_explainer | Rationale / Explanation | 최종 판정 근거, 핵심 evidence, operator 영향 설명 생성 | 입력: snapshot, rationale graph 출력: explanation text, rationale subgraph |
고도화 | XAI/공공 실무 중요 |
| claim_workspace_ui | Productization Layer | claim, evidence, stance, uncertainty를 분석관이 검토하는 운영 화면 제공 | 입력: graph/snapshot/API 출력: 검토 UI, 감사 로그 |
고도화 | 실증/시연용 핵심 |
| causal_bridge_interface | Epistemic-to-Causal Interface | epistemic snapshot을 확산·영향 예측용 feature vector로 변환 | 입력: belief, uncertainty, contradiction ratio 등 출력: causal input vector |
장기 | 연구·운영 연결층 |
| propagation_modeler | Causal Graph / Risk Modeling | 확산·노출·행동 변화 인과 구조를 추정하고 리스크 드라이버 식별 | 입력: causal input vector, time-series 출력: causal DAG, risk drivers |
장기 | 데이터 축적 필요 |
| intervention_simulator | Intervention & Simulation | fact-check, label, counter-message 등의 개입 효과를 do-연산 관점에서 시뮬레이션 | 입력: causal DAG, action plan 출력: expected_risk_delta, CI, recommended_actions |
장기 | 2차년도 후반 적합 |