---
features:
  - name: input
    dtype: string
  - name: output
    dtype: json
  - name: schema
    list:
      - name: label
        dtype: string
      - name: description
        dtype: string
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/maven_arg_train.jsonl
      - split: validation
        path: data/maven_arg_validation.jsonl
---

# MAVEN-Arg (mneb event-extraction format)

MAVEN-Arg (Wang et al., ACL 2024) converted into the mneb `json_structures`
event-extraction format. It adds document-level, role-typed arguments to MAVEN events over
English Wikipedia documents. This is an independent document-level dataset intended for
`mneb/maven-arg`; it does not replace or modify the sentence-level, detection-only
`mneb/maven` dataset.

One record is one complete source document. Character offsets are end-exclusive and satisfy
`input[start:end] == text`. Event-type and argument-role strings are retained verbatim.
Every trigger mention in a source event cluster becomes an independent event entry and
inherits that cluster's full argument annotation. Events with no arguments are retained.

## Splits and counts

| Split | Records | Source event clusters | Event entries | Source argument objects | Expanded argument spans | Event types |
|---|---:|---:|---:|---:|---:|---:|
| `train` | 2,913 | 64,923 | 70,775 | 190,479 | 380,830 | 162 |
| `validation` | 710 | 15,556 | 16,996 | 46,458 | 85,890 | 161 |
| **total** | **3,623** | **80,479** | **87,771** | **236,937** | **466,720** | **162** |

`validation` is the official raw `valid` split, renamed for consistency. The official test
split is excluded because it contains candidate event mentions but not the complete gold
event clusters and argument annotations required for this conversion.

## Argument expansion

MAVEN-Arg attaches arguments to event clusters, not individual trigger mentions. Therefore,
each emitted trigger entry receives a copy of every argument on its source cluster.

- A direct `{content, offset}` argument becomes one `{role, text, start, end}` span.
- An `{entity_id}` argument is resolved against the document's entity coreference clusters
  and expands to every mention span in the referenced cluster, each with the same role.
- Expansion preserves coreferent mention spans as event arguments, but does not encode the
  fact that those mentions corefer.

All 94,922 entity references across train and validation resolve successfully. All source
trigger, direct-argument, and entity-mention offsets were checked against the original
document text; no annotation was silently dropped.

## Deliberately omitted layers

This release contains event extraction only: trigger detection/classification plus event
argument extraction. It does not emit `entities` or `relations`, and in particular does not
emit entity or event coreference links. Source document/event/entity/mention ids, numeric
`type_id`, and `negative_triggers` are also omitted.

Entity and event coreference are outside the scope of this event-extraction release; no
cluster-to-pair conversion or coreference relation layer is planned here.

## Files

- `data/maven_arg_train.jsonl` and `data/maven_arg_validation.jsonl`: converted data.
- `maven_arg_label_role_summary.md`: event-label and argument-role counts.
- `maven_arg_vis.html`: deterministic sampled visualizer covering every event type in each
  split where that type exists.
- Converter: `scripts/preprocess_maven_arg.py` in the source workspace.

## Provenance

Source dataset: THU-KEG/MAVEN-Argument, accompanying *MAVEN-Arg: Completing the Puzzle of
All-in-One Event Understanding Dataset with Event Argument Annotation* (ACL 2024). The
documents originate from English Wikipedia. Users should review the upstream repository and
dataset terms before redistribution or use.