mneb/genia2011

GENIA 2011 (mneb format) — joint NER + nested event extraction

The BioNLP Shared Task 2011 GE corpus (Kim et al., 2011) converted into the mneb joint entities + json_structures format. GENIA is a biomedical entity and end-to-end event extraction dataset over PubMed abstracts and PMC full texts.

This dataset keeps event-as-argument nesting. Roughly a third of GENIA's argument links point at another event rather than an entity — regulation events take other events as their Theme or Cause — and prior conversions (including TextEE's) drop them. Here they are kept without leaving the flat mneb record shape: every argument is a plain {role,text,start,end} span, and an argument that points at an event carries that event's trigger span. See Event arguments.

Char offsets are character-based and end-exclusive (input[start:end] == text). One record = one source .txt file (PMC full-text sections are not merged into papers).

Splits

Split	Records	With events	Events
train	908	765	9,560
validation	259	225	2,988
Total	1,167	990	12,548

There is no labelled test split. The official blind-test archive contains 347 .txt/.a1 inputs, but BioNLP-ST did not publicly release their .a2 event gold. This release therefore contains the official train/devel partition only; treating the blind inputs as empty output would create false negatives. Note that this differs from TextEE, which discards the official boundary and makes five random re-splits instead — no number reported on a TextEE split is directly comparable to this one.

Supported tasks and retained layers

• NER: protein and generic entity mentions in output.entities.
• Event detection (ED): event triggers in output.json_structures.
• Event argument extraction (EAE) and end-to-end event extraction (E2E): trigger and argument spans, including span-linked event arguments.
• Joint NER + EE: both layers occur in the same record and use the same character offsets.

Equivalence, event attributes and normalisations are not exposed, so this release does not claim relation extraction, coreference, negation or speculation as supported tasks.

Entity layer — 2 types, 16,976 mentions

Entity type	Train	Validation	Total
entity	480	181	661
protein	11,625	4,690	16,315
Total	12,105	4,871	16,976

protein is the public form of .a1 label Protein; entity is the public form of the generic .a2 label Entity used for non-protein physical participants, sites and locations. The names are uniformly lower-case with spaces, matching mneb's public entity-label policy.

Record format

{
  "input":  "<document text>",
  "output": {
    "entities": {"protein": [<span>, ...], "entity": [<span>, ...]},
    "json_structures": {"<event type>": [<event>, ...]}
  },
  "schema": {"entities": [<entity definition>, ...],
             "json_structures": [<event definition>, ...]}
}

The active data/train.jsonl and data/validation.jsonl use lower-case, space-separated event labels (for example gene expression), preserving the convention introduced by yangwang825. Compatibility files prefixed genia2011_ retain the official event strings (for example Gene_expression). NER labels are normalized in both views.

An event is {"trigger": {"text","start","end"}, "arguments": [<arg>, ...]}. There is nothing else: no type field on the event (its type is the json_structures key), and no object nested inside an argument.

Event arguments

Every argument has the same four keys, whether it points at an entity or at another event:

{"role": "Theme", "text": "interferon regulatory factor 4", "start": 19, "end": 49}
{"role": "Theme", "text": "expression",                     "start": 55, "end": 65}

The first is an entity mention. The second is an event link: (55, 65) is the trigger span of a gene expression event, which is listed at the top level of the same record. This is how mneb expresses links generally — repeat the span, no ids (cf. mneb/bc5cdr, whose relation head/tail repeat the entity spans).

Because a linked child must be reachable, every event appears at the top level, not only the roots. To resolve links:

def resolve(js):
    """Index every event by its trigger span, then read arguments as links where they match."""
    by_span = {}
    for etype, evs in js.items():
        for ev in evs:
            by_span.setdefault((ev["trigger"]["start"], ev["trigger"]["end"]), []).append((etype, ev))
    for etype, evs in js.items():
        for ev in evs:
            for a in ev["arguments"]:
                target = by_span.get((a["start"], a["end"]))      # None => entity mention
                yield etype, ev, a, target

A real example — "Down-regulation of interferon regulatory factor 4 gene expression…". Two top-level events; the Theme link Down-regulation → expression is the nesting:

"negative regulation": [{
  "trigger": {"text": "Down-regulation", "start": 0, "end": 15},
  "arguments": [{"role": "Theme", "text": "expression", "start": 55, "end": 65}]}],
"gene expression": [{
  "trigger": {"text": "expression", "start": 55, "end": 65},
  "arguments": [{"role": "Theme", "text": "interferon regulatory factor 4",
                 "start": 19, "end": 49}]}]

How faithful the span links are

• Telling a link from an entity mention: just 1 of the 11,590 entity-valued arguments sits on a span that is also a trigger, so the test "this argument's span matches a trigger span" has a single false positive corpus-wide.
• Telling which event a link points at: 3,860 of the 5,502 links (70.2%) match exactly one event of the right type. The other 1,642 (29.8%) land on a trigger span shared by several same-type events, and the span cannot disambiguate them — GENIA 2011 is the worst of the three BioNLP corpora here, because one regulation trigger routinely heads several events.
• Consequently 1,012 of the 13,560 raw E lines (7.5%) come out byte-identical to another entry of the same type and are collapsed, leaving 12,548 events. Those are exactly the parents that differed only in an unresolvable choice of child; keeping both copies would double-count in any set-based metric.

Everything else round-trips: the offset invariant holds on every span, and the set of emitted (type, trigger, role/span) signatures equals the same set computed straight off the raw standoff, for every document.

Statistics

• 9 event types, 10 role types (role strings kept verbatim, so Theme2/Theme3/ Theme4/Site2 are not collapsed into their base role).
• 17,092 raw argument links = 11,590 entity-valued + 5,502 event-valued (32.2% of all argument links are event-to-event). After the collapse above the files hold 15,652 argument instances = 11,354 entity spans + 4,298 span links.
• 4,964 events (36.6%) take at least one event argument.
• Raw nesting depth histogram {1: 8596, 2: 4209, 3: 710, 4: 44, 5: 1} — max depth 5.
• Nesting is driven entirely by the three regulation types; no other event type ever takes an event argument, and only Theme and Cause are ever event-linked.

Full type/role inventory and nesting patterns: genia2011_label_summary.md. Browsable rendering: genia2011_vis.html (open directly; data embedded, no server needed).

How this was derived

Required by clause 5 of the source licence. Built from the official BioNLP-ST_2011_genia_train_data_rev1 and ..._devel_data_rev1 archives:

1. Each document's .a1 (Protein entities) and .a2 (event triggers, Entity spans and E event lines) are parsed into a single text-bound annotation map; their id spaces do not collide.
2. Protein and Entity annotations become NER mentions under normalized public labels.
3. Each E line becomes an event grouped under its own type; Role:T… arguments become the entity's span and Role:E… arguments become the child event's trigger span. Both come out in the same {role,text,start,end} shape.
4. Every event is listed at the top level, so a linked child is always resolvable. Entries that are byte-identical under one event type are then collapsed.
5. Equivalence (*), attribute (A/M: Negation, Speculation) and normalisation (N) lines are not carried over. The converter and compatibility files preserve official event-type strings; the active files normalize only event-type labels. Roles remain verbatim in both views.

The raw standoff gives labels but no prose label descriptions. The descriptions embedded in the full schema and published in schema.json are concise mneb-authored paraphrases based on the official task definition and paper; they are not source annotations or quotations.

Verification built into the converter: every emitted span is re-checked against the source text (input[start:end] == text, 0 failures); the set of emitted (type, trigger, role/span) signatures is compared against the same set computed straight off the raw standoff and is equal for every document; and a link audit reports, for every event-valued argument, whether its child is uniquely identifiable from the span (the numbers quoted above).

Licence and terms of use

Distributed under the BioNLP Shared Task 2011 licence terms, which permit redistribution with these conditions:

• Research use. The annotations are to be used primarily for scholarly research in NLP, IE and related disciplines. Derived data may be produced only for bona fide research of a non-profit nature.
• Commercial or non-academic public use requires permission from the BioNLP-ST'11 organisers.
• Abstracts come from PubMed® (U.S. National Library of Medicine) and are subject to the PubMed licence terms. Full texts come from the PMC Open Access Subset; each article carries its own Creative Commons or similar licence.
• Copyright in the annotations belongs to the BioNLP-ST'11 organisers.

Please read the full terms at the source link above before redistributing or building on this data.

Citation

@inproceedings{Kim11genia2011,
  author    = {Jin-Dong Kim and Yue Wang and Toshihisa Takagi and Akinori Yonezawa},
  title     = {Overview of Genia Event Task in BioNLP Shared Task 2011},
  booktitle = {Proceedings of BioNLP Shared Task 2011 Workshop},
  year      = {2011}
}