---
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_train.jsonl
      - split: validation
        path: data/maven_validation.jsonl
---

# MAVEN (mneb/casie format)

MAVEN (Wang et al., EMNLP 2020) converted into the mneb `json_structures` event-extraction
format. MAVEN is a massive general-domain event **detection** dataset — 168 event types over
English Wikipedia documents. It is ED-only (no argument annotation), so each record's `output`
carries only `json_structures`; every event's `arguments` is always `[]`. One record = one
sentence. ALL sentences are kept, including event-less ones (`json_structures: {}`). Char
offsets are end-exclusive (`input[start:end] == text`). Event-type strings are kept verbatim
from the raw release.

## Splits

Official MAVEN document split:

| Split | Docs | Sentences | Event mentions |
|---|---:|---:|---:|
| `train` | 2,913 | 32,431 | 77,993 |
| `validation` | 710 | 8,042 | 18,904 |

`validation` is MAVEN's raw `valid` split, renamed for naming consistency with the other
converted datasets. The official `test` split is **not included**: its annotations are withheld
for the CodaLab leaderboard, so it carries no gold event type or mention, only candidate
triggers.

## What is dropped from the raw release

- The unlabeled `test` split.
- `negative_triggers` and other candidate lists (explicitly-labeled non-events).
- Event coreference: MAVEN groups multiple mentions under one event instance. The flat
  per-sentence format used here cannot represent that grouping, so mentions become independent
  trigger entries — mention counts are preserved exactly (nothing deduplicated or merged).
- Numeric `type_id`s, mention/event ids, and the raw per-sentence token lists.

## Token -> char note

MAVEN's raw trigger spans are token indices into per-sentence token lists that are lossy
(lowercased, rare words replaced by the literal string `unk`, PTB escapes for brackets/quotes).
Triggers were re-anchored to the raw sentence text via greedy earliest-match case-insensitive
alignment, trying PTB-variant candidates alongside the raw token, plus a neighbor-window
fallback search for `unk`-substituted tokens. This recovered 100% of mentions: 0 drops, 0
offset failures across both splits. The emitted trigger `text` is always the authoritative
sentence slice (`input[start:end]`); it differs from the raw `trigger_word` in 795 (train) + 200
(validation) cases, all case-only.

## Provenance & license

Source: THU-KEG/MAVEN-dataset (data distributed via Google Drive / Tsinghua Cloud), MIT
license. Document text is drawn from English Wikipedia (CC BY-SA). The event schema is derived
from FrameNet. Annotations are crowd-sourced.

See `maven_label_summary.md` for the full 168-type inventory and per-split counts, and
`maven_vis.html` for a sampled visualiser (stratified sample, ≤20 records/type per split + 10
empty records per split).