MLEE (mneb format) — joint NER + nested event extraction

The Multi-Level Event Extraction corpus (Pyysalo et al., Bioinformatics 2012) converted into the mneb joint entities + json_structures format. MLEE annotates biomedical entities and events across multiple levels of biological organisation — from molecular through cellular and tissue to organism level — over PubMed abstracts on angiogenesis.

This dataset keeps event-as-argument nesting. A third of MLEE's argument links point at another event rather than an entity, 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 MLEE document.

Splits

The official MLEE partition is preserved.

Split	Records	With events	Events
train	131	131	3,206
validation	44	44	1,102
test	87	87	2,132
Total	262	262	6,440

TextEE instead discards the official boundary and makes five random re-splits, so no number reported on a TextEE split is directly comparable to this one.

Supported tasks and retained layers

• NER: all 16 source entity types 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.

The source also contains relation, equivalence and event-attribute layers. They are not exposed in this release, so it should not be treated as an RE, coreference, negation or speculation benchmark.

Entity layer — 16 types, 8,290 mentions

Entity type	Train	Validation	Test	Total
anatomical system	9	1	8	18
cell	714	152	332	1,198
cellular component	77	28	40	145
developing anatomical structure	3	1	2	6
dna domain or region	31	4	0	35
drug or compound	435	202	307	944
gene or gene product	1,460	500	1,001	2,961
immaterial anatomical entity	8	3	4	15
multi-tissue structure	259	89	166	514
organ	82	41	53	176
organism	359	126	237	722
organism subdivision	20	7	22	49
organism substance	56	26	60	142
pathological formation	382	171	357	910
protein domain or region	17	10	2	29
tissue	234	70	122	426
Total	4,146	1,431	2,713	8,290

Source labels such as Gene_or_gene_product are exposed as lower-case, space-separated public labels such as gene or gene product. Exact duplicate annotations with the same label, span and text are collapsed (one duplicate corpus-wide); a span annotated with more than one type is retained once under each type.

Record format

{
  "input":  "<document text>",
  "output": {
    "entities": {"<entity_type>": [{"text":"...","start":0,"end":4}]},
    "json_structures": {"<event_type>": [<event>, ...]}
  },
  "schema": {"entities": [<entity definition>, ...],
             "json_structures": [<event definition>, ...]}
}

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": "Cause",  "text": "anti-VEGF neutralizing antibody", "start": 1705, "end": 1736}
{"role": "Theme",  "text": "stimulated",                     "start": 1611, "end": 1621}

The first is an entity mention. The second is an event link: (1611, 1621) is the trigger span of a Positive_regulation 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 from PMID-10586954 — "The insulin-conditioned RPE cell media stimulated capillary endothelial cell proliferation, an effect that was completely blocked by anti-VEGF neutralizing antibody". Three top-level events; the Theme chain blocked → stimulated → proliferation is the nesting:

"Negative_regulation": [{
  "trigger": {"text": "blocked", "start": 1694, "end": 1701},
  "arguments": [{"role": "Theme", "text": "stimulated", "start": 1611, "end": 1621},
                {"role": "Cause", "text": "anti-VEGF neutralizing antibody",
                 "start": 1705, "end": 1736}]}],
"Positive_regulation": [{
  "trigger": {"text": "stimulated", "start": 1611, "end": 1621},
  "arguments": [{"role": "Theme", "text": "proliferation", "start": 1649, "end": 1662}]}],
"Cell_proliferation": [{
  "trigger": {"text": "proliferation", "start": 1649, "end": 1662},
  "arguments": [{"role": "Theme", "text": "capillary endothelial cell",
                 "start": 1622, "end": 1648}]}]

A conversion that simply dropped event-valued arguments would say only that something was blocked by an antibody, losing that what was blocked is the stimulation of proliferation.

How faithful the span links are

• Telling a link from an entity mention: on MLEE, 0 of the 5,767 entity-valued arguments sit on a span that is also a trigger. The test "this argument's span matches a trigger span" therefore has no false positives here.
• Telling which event a link points at: 2,417 of the 2,832 links (85.3%) match exactly one event of the right type. The other 415 (14.7%) land on a trigger span shared by several same-type events, and the span cannot disambiguate them.
• Consequently 238 of the 6,678 raw E lines (3.6%) come out byte-identical to another entry of the same type and are collapsed, leaving 6,440 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

• 29 event types, 14 role types (role strings kept verbatim, so Theme2, Participant2..4 and Instrument2 are not collapsed into their base role).
• 8,599 raw argument links = 5,767 entity-valued + 2,832 event-valued (32.9% of all argument links are event-to-event). After the collapse above the files hold 8,155 argument instances = 5,660 entity spans + 2,495 span links.
• 2,260 events (33.9%) take at least one event argument.
• Raw nesting depth histogram {1: 4416, 2: 1981, 3: 269, 4: 10} — max depth 4.
• Nesting is almost entirely driven by the three regulation types; the sole exception is Planned_process, which takes an event argument 9 times. Only Theme and Cause are ever event-linked.

Note on event counts. Two of the 6,678 raw E lines are byte-identical duplicate annotations (PMID-16076702 E28/E29 and PMID-19540587 E11/E25 — same type, same trigger, same arguments); they collapse under the same rule as everything else.

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

How this was derived

Built from the MLEE-1.0.2-rev1 standoff release (standoff/full/*.{txt,ann}), with the official split membership taken from the filenames in standoff/{development/train, development/test, test/test}:

1. Each document's .ann (a1 + a2 merged: entity mentions, event triggers and E event lines) is parsed into a single text-bound annotation map.
2. Every non-trigger text-bound annotation becomes an NER mention. Source entity labels are rewritten to lower-case words separated by spaces; exact duplicates are collapsed while same-span, different-type annotations remain separate.
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. Relation (R), equivalence (*) and attribute (A/M: Negation, Speculation) lines are not carried over. Event-type and role strings are kept verbatim.

The standoff source provides annotation labels, not prose label definitions. The descriptions in the embedded schema and schema.json are concise mneb-authored paraphrases of the official MLEE paper/site and the referenced biomedical ontologies; they are documentation, not text copied from the raw annotation files.

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

• Annotations are licensed under Creative Commons BY-NC-SA 3.0. This is a non-commercial, share-alike licence: derivative works must carry the same terms, and commercial use is not permitted. Please attribute by citing the paper below and linking to http://www.nactem.ac.uk/MLEE/.
• Abstracts are from PubMed, a database of the U.S. National Library of Medicine; see the NLM copyright information.

Citation

@article{Pyysalo12mlee,
  author  = {Sampo Pyysalo and Tomoko Ohta and Makoto Miwa and Han-Cheol Cho and
             Jun'ichi Tsujii and Sophia Ananiadou},
  title   = {Event extraction across multiple levels of biological organization},
  journal = {Bioinformatics},
  volume  = {28},
  number  = {18},
  pages   = {i575--i581},
  year    = {2012}
}