Event Extraction

Bonan Min
[email protected]
Some slides are based on class materials from Ralph Grishman
 Event Extraction: An Example

Rainfall in July continued ... in Ethiopia, causing displacement…

Timex: 2017-07 Geo-Political Entity

Preprocessing
◦ Tagging named entities, mentions and value mentions (e.g., time)

2
 Event Extraction: An Example

Rainfall in July continued ... in Ethiopia, causing displacement…

E1: Weather E2: Displacement
Timex: 2017-07 Geo-Political Entity

Preprocessing
◦ Tagging named entities, mentions and value mentions (e.g., time)

Event Extraction
◦ Event detection: detect and classify event mentions

3
 Event Extraction: An Example
Place
Time

Rainfall in July continued ... in Ethiopia, causing displacement…

E1: Weather E2: Displacement
Timex: 2017-07 Geo-Political Entity

Preprocessing
◦ Tagging named entities, mentions and value mentions (e.g., time)

Event Extraction
◦ Event detection: detect and classify event mentions
◦ Argument Extraction: attach event arguments Who, When (Time), and Where (Place)

4
 Event Extraction: An Example
Place
Time

Rainfall in July continued ... in Ethiopia, causing displacement…

E1: Weather E2: Displacement
Timex: 2017-07 Geo-Political Entity

Preprocessing
◦ Tagging named entities, mentions and value mentions (e.g., time)

Event Extraction
◦ Event detection: detect and classify event mentions
◦ Argument Extraction: attach event arguments Who, When (Time), and Where (Place)

5
Event Extraction
Scenario Template (MUC: Message Understanding Conference)
◦ The scenario template task originally was the IE task for the MUC
evaluations
◦ Identify participants, locations, dates etc. of a class of events -- a naval engagement, a
terrorist incident, a joint venture.
◦ A single template included related information, such as an attack and its effects; this led to
some relatively complex templates
◦ With later MUCs (6 and 7), the task narrowed to single events or closely
related events -- executive succession, rocket launchings

For the ACE evaluations, this became the event extraction task.
An event is
◦ a specific occurrence involving participants.
◦ something that happens.
This is broadly defined, so we
◦ frequently described as a change of state. need an inventory of types
 ACE Events
Event Type Event Subtype
Life Be-Born, Marry, Divorce, Injure, Die
Movement Transport
Transaction Transfer-Ownership, Transfer-Money
Business Start-Org, Merge-Org, Declare-Bankruptcy, End-Org
Conflict Attack, Demonstrate
Contact Meet, Phone-Write
Personnel Start-Position, End-Position, Nominate, Elect
Justice Arrest-Jail, Release-Parole, Trial-Hearing, Charge-Indict, Sue, Convict,
Sentence, Fine, Execute, Extradite, Acquit, Appeal, Pardon

S1: Six murders occurred in France, including the assassination of Bob and the killing of Joe.
S2. Six men were murdered, including Bob (in Paris) and Joe (in Reims).
Conflict.Attack Life.Die
 Event Anchors/Triggers
Many Events anchor on a single verb or noun
Six murders occurred in France, including the
assassination of Bob and the killing of Joe.

There are some cases where multiple words could reasonably be

called the trigger
Foo Corp. had previously filed Chapter 11 in 2001.
The leaders held a meeting in Beijing.
The company was ordered to pay a fine of $300,000.
◦ We can choose a word as the anchor (and do so consistently*)

* Details are in the ACE guideline:

https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/english-events-guidelines-v5.4.3.pdf
 Event Arguments: Participants (ACE)
Most Event arguments will be participants in the Event
◦ Taggable Entities that are involved in the Event

Event type Event subtype Argument role types*

Life Be-Born, Marry, Divorce Person
ACE argument Injure, Die Agent, Victim, Instrument
role types (3 Movement Transport Agent, Artifact, Vehicle,
of 33 subtypes Origin, Destination
are shown)
Transaction Transfer-Ownership Buyer, Seller, Beneficiary,
Artifact, Price

An alternative set
Transaction.Transfer-Ownership
(in CAMEO):
[His brother] bought [him] [a new car] with [$20,000]. • Agent
Buyer Beneficiary Artifact Price • Patient
• Location
*All event types have Place and Time arguments. • Time
Event Arguments: Attributes (ACE)
These argument slots can be filled by certain Values identified within the
scope of the Event

Event-specific attributes

General event attributes

Examples
Clinton also touched on the matter of American Edmond Pope who is being
tried in a closed court in Russia on charges of [spying]
TRIAL-HEARING Crime-Arg

Greg Lashutka was elected [mayor of Columbus] in 1993.

ELECT Position-Arg

46-year-old Abu Talib was sentenced to [life imprisonment] in 1990 in Sweden

for terrorist acts in Amsterdam. Sentence-Arg
SENTENCE

We were married in [Spain] [two years ago].

MARRY Place-Arg Time-Arg
 Event and sub-Events
S1: Six murders occurred in France, including the assassination of Bob and the killing of Joe.
Anchor Arguments
role Mention/Value
murders location France
assassination Victim Bob
location France
killing Victim Joe
location France

S2. Six men were murdered, including Bob (in Paris) and Joe (in Reims).
Anchor Arguments
role Mention/Value
murdered Victim six men, Bob, Joe
location Paris, Reims
Event Properties: Polarity and Genericity
Polarity
◦ An Event is NEGATIVE when it is explicitly indicated that the Event did not
occur. All other Events are POSITIVE.
His wife was sitting on the backseat and was not hurt.
They backed out of the purchase at the last minute.

Genericity
◦ An Event is SPECIFIC if it is understood as a singular occurrence at a
particular place and time, or a finite set of such occurrences. All other Events
are GENERIC.

Salat Hassen called on countries that give aid.

The group specialized in transporting illegal weapons.
Event Properties: Tense
Tense is determined with respect to the speaker or author. We
will refer to the time of publication or broadcast as the textual
anchor time.
◦ PAST is used for those Events that occur prior to the textual anchor time
He traveled to Houston in late September.

◦ FUTURE is used for those Events that have not yet occurred at the textual anchor
time.
He plans to meet with lawmakers from both parties.
When he's born, he'll be named after his father.

◦ PRESENT is used for those Events that occur at the textual anchor time
The airline is in the midst of a major aircraft purchase from Airbus Industries.
 Event Properties: Modality
An Event is ASSERTED when the author or speaker makes reference to it as
though it were a real occurrence
He traveled to Houston in late September.

All other Events will be annotated as OTHER:

Types Examples
Believed Rumors of arrests circulated in Vancouver.
Hypothetical Should he not pay the money, they would kill him.
Commanded/Re He was ordered to return to Moscow.
quested He asked the United States to give money to his country.
Threatened/Pro The mayor's accomplices had threatened to kill Mr. Tatum if he refused.
posed/Discusse He thought about paying up. The United Nations has warned their people not to
d Events take the ferry.
Desired Events They wanted to acquire the company last year.
Promised He said he would leave town.
Otherwises He decided it would probably be a good idea to meet with him.
Event Coreference
When two Event mentions refer to the same Event, then they
corefer
Ukraine International Airlines Flight 752 crashed Wednesday after takeoff
from Tehran's airport. The crash came hours after Iran fired missiles...

What about when one mention refers only to a part of the

other? What about modality?
Three people have been convicted in the operation, including Smith and Jones. …
Smith and Jones were found guilty of selling guns to straw purchasers.

Maddux was killed in Philadelphia. …. Einhorn is accused of killing Maddux.

... the Boston Marathon terror attack. On April 15, 2013, double bombings near
the finish line of the Boston Marathon killed three people and injured at least 264.
... The bombs exploded 12 seconds apart near the marathon's finish line on
Boylston Street.
Event Extraction Pipeline (Stages)
Anchor Argument Property Event
identification identification assignment coreference

Anchor identification
◦ finding event anchors(the basis for event mentions) in text and assigning them
an event type

Argument identification
◦ determining which entity mentions, timexes, and values are arguments of each
event mention

Property assignment
◦ determining the values of the modality, polarity, genericity, and tense
properties for each event mention

Event coreference
◦ determining which event mentions refer to the same even
David Ahn. 2006. The stages of event extraction
Event Extraction Pipeline (Stages)
Anchor Argument Property Event
identification identification assignment coreference

Each stage can we modeled as a classification problem

◦ SVM, MaxEnt, CNN/MLP, etc

To understand what works, let’s start with extraction templates/patterns:

Text Pattern Type

President Trump attacked POLITICIAN/PRESIDENT attacked Verbal Attack

democrats … ORG/POLITICAL_PARTY
A rocket attacked on Taji Air WEAPN attacked on FAC/MILITARY_BASE Physical/Material
Base … Attack
hackers laid a cyber attack ... HACKERs nsub_of-1 laid dobj attack Cyber Attack
Features For Event Anchors
The attack killed 7 and injured 20.
Hamas launched an attack.

Lexical features: full word, lowercase word, lemmatized word, POS tag,
depth of word in parse tree
WordNet features: Synset for words in categories such as noun, verb,
adjective, and adverbs
Left context (e.g., 3 words): lowercase, POS tag
Right context (e.g., 3 words): lowercase, POS tag
Dependency features: labels of the dependency relations, dependent/head
words, or both. An example is <*, subj, killed>
Related entity features: Hamas attacked Israeli army
◦ Entity types, constituent head words, etc
Trump attacked CNN
◦ Often used in conjunction with dependency relations
Features For Argument Identification
Hamas launched an attack.

Anchor word of event mention: full, lower-case, POS tag, and depth in parse
tree
Event type of event mention, e.g., Conflict.Attack
Constituent head word of entity mention: full, lowercase, POS tag, and
depth in parsetree
Entity type and mention type (name, pro-noun, other NP) of entity mention
Dependency path between anchor word and constituent head word of
entity mention
◦ Often expressed as a sequence of labels, of words, and/or of POS tags
◦ Often used in conjunction with entity types to reduce sparsity
◦ E.g., ORG.Military subj launched obj attack
Features For Assigning Attributes
Train a separate classifier for each attribute
◦ Genericity, modality, and polarity are each binary classification tasks
◦ Tense is a multi-class task.

Similar features for the anchor identification task, with the exception of
the lemmatized anchor word and the WordNet features

Hamas attacked Israeli army

Trump attacked CNN
Features For Coreference Classification
Ukraine International Airlines Flight 752 crashed Wednesday after takeoff
from Tehran's airport. The crash came hours after Iran fired missiles...

Candidate anchor + anaphor anchor, also POS tag and lowercase

Candidate event type + anaphor event type
◦ Event mentions of the same types are more likely to refer to the same event

Distance between candidate and anchor

◦ measured in sentences
◦ Nearby mention are more likely to to refer to the same event
Number, heads, and roles of shared arguments
◦ Shared arguments: same entity/timex/value with same role

Number, heads, and roles of candidate arguments that are not anaphor arguments
Number, heads, and roles of anaphor arguments that are not candidate arguments
Heads and roles of arguments shared by candidate and anaphor in different roles
Candidate modality value + anaphor modality value
◦ also for polarity, genericity, and tense
 Convolutional Neural Networks
for Event Extraction
CNN for event classification and argument identification
◦ WE: word embeddings
◦ PF: position embeddings features
◦ EF: event embeddings features
◦ Local features: embeddings of trigger, argument, and words around them (e.g. left 3 words, right 3
words). These local features are concatenated with the max-pool to form the final input features to the
dense network, for prediction
Word embeddings Word-level features CNN filters Max-pool

Generalization Vector NN
concatenation output
over words
focuses
Avoid (extensive)
on feature engineering
relief
and
recovery
activities
in
the
non-conflict
affected
WE PF EF
states WE: word embeddings
PF: position embeddings features
,
EF: event embeddings features

Trigger Argument

Local features around trigger and argument

Chan et al., Rapid Customization for Event Extraction. ACL 2019.

Dynamic Multi-Pooling CNN
For argument classification, each of the 3 sections (before the argument cameraman, after the trigger
fired, and in between these two) may be indispensable
◦ Traditional CNN take a single max from each feature maps

DMCNN for the argument classification stage

◦ splits each feature map into three parts according to the candidate argument and predicted trigger
◦ Instead of using one max value for an entire feature map to represent the sentence, keeps the max value of each
split part

Chen et al., Event Extraction via Dynamic Multi-Pooling Convolutional Neural Networks. ACL-IJCNLP 2015.
Problems with Pipeline Models
Anchor Argument Property Event
identification identification assignment coreference

Problem 1: Error propagation

◦ Trigger/anchor ID and arugment ID are separate stages; can’t recover error

Problem 2: Steps are decoupled for events & arguments in a sentence

◦ The Pipeline model can detect cameraman as argument to died (because they are
nearby).

◦ cameraman is far away from fired, it may fail to recognize cameraman as its Target
argument with the local features.
◦ Can we generate global features to encode that a Victim argument for the Die event is
often the Target argument for the Attack event in the same sentence?
Joint Trigger and Argument Extraction via RNN

Model Trigger Argument

Feature-based Model (separate) (Li et al. 2013)† 65.9 43.9

Feature-based Model (joint) (Li et al. 2013)† 67.5 52.7

Deep Learning Model (separate) (Chen et al. 2015)‡ 69.1 53.5

Deep Learning Model (joint) (Nguyen et al. 2016) 69.3 55.4

Thien Huu Nguyen, Kyunghyun Cho and Ralph Grishman. Joint Event Extraction via Recurrent Neural Networks. NAACL-HLT 2016.