ARGUS: Role-Aware Event Knowledge Graphs for U.S. Employment-Discrimination Complaints

The authors present ARGUS, a source-grounded, role-aware pipeline for constructing document-level Event Knowledge Graphs (EKGs) from U.S. employment-discrimination complaints, representing events with participants, temporal relations, and causal relations.

Source: [9]

The document-scoped FAQ benchmark contains 300 questions over 10 CourtListener complaints, with 30 questions per document. Because each question is paired with its source complaint, the experiment measures representation and reasoning after the relevant document is known, not first-stage retrieval across CourtListener.

Source: [15], [16]

The corpus contains 153 CourtListener U.S. federal employment-discrimination complaints from five district courts. These are initiating pleadings describing allegations, not facts established by a court; accordingly, ARGUS events represent complaint allegations and should not be interpreted as adjudicated findings.

Source: [13]

For claim-type classification, the paper uses 130 complaints with 176 label assignments over 23 concepts, with LLM-generated silver labels reviewed by humans. The reported held-out classification test set contains 22 documents.

Source: [6], [7], [25]

The document-scoped FAQ compares raw-document retrieval, hybrid text plus EKG, and EKG-only retrieval while holding GPT-4o and the answering prompt fixed; answer quality is measured using Token-F1 against the gold answer.

Source: [2], [14], [15], [17]

The paper compares raw complaint text encoded from a 512-token input, a linearized entity/event EKG, and a graph-structured EKG using participation, temporal, and causal edges over the same classification label space.

Source: [7]

The authors characterize ARGUS as a research prototype for retrieval and organization rather than legal advice, and state that automatically extracted roles, events, and causal relations may be incomplete or wrong and should not be treated as legal findings or used without attorney review.

Source: [22]

ARGUS processes complaint FACTS narratives through sentence-level fact extraction, semantic chunking, two-stage mini-graph extraction, and deterministic plus LLM-assisted document-level merging; the authors evaluate components separately to distinguish errors from fact selection, event and relation extraction, and graph merging.

Source: [10]

The event representation uses entity records with canonical legal roles and event records with event type, participants, temporal information, source evidence, and confidence. It supports BEFORE, AFTER, OVERLAP, and SAME-TIME temporal relations and CAUSES, ENABLES, and PREVENTS causal relations, retaining relations only when supported by source text.

Source: [20]

In two-stage mini-graph extraction, Stage 1 extracts grounded entities and events without temporal or causal edges; Stage 2 receives the original chunk and Stage 1 output and adds only source-supported temporal and causal relations. Both stages require strict JSON, preservation of sentence identifiers, and no unsupported entities, events, dates, or relations.

Source: [1], [26]

The paper asks whether the staged pipeline can produce reliable, source-grounded legal event graphs; whether those graphs improve legal classification and document-scoped question answering over text-based representations; and whether benefits extend to external and open-retrieval legal QA.

Source: [9], [24]

For the reported graph classifier, LEGAL-BERT encodes graph-node descriptions; two mean-neighbor graph-convolution layers are followed by mean pooling and a 23-label classification head. The reported configuration jointly fine-tunes LEGAL-BERT and graph layers using weighted binary cross-entropy and AdamW.

Source: [4]