Lei Sha, Jing Liu, Chin-Yew Lin, Sujian Li, Baobao Chang, Zhifang Sui RBPB : Regularization-Based Pattern Balancing Method for Event Extraction Lei Sha, Jing Liu, Chin-Yew Lin, Sujian Li, Baobao Chang, Zhifang Sui
Event extraction Identifying the trigger of an event, Identifying the arguments of the event Distinguishing the arguments’ corresponding roles
ACE event extraction task
Steps 1. Extract Trigger 2. Identify Arguments 3. Classify Roles 1. Identify Trigger 2. Classify Trigger 3. Identify Arguments 4. Classify Arguments 1. Classify Trigger 2. Classify Roles
Related work Pattern based methods for identifying event type (Kim and Moldovan, 1993; Grishman et al.,2005; Ji and Grishman, 2008;) Bootstrap for more patterns (Huang and Riloff, 2012; Liu and Strzalkowski, 2012) Feature-based classification methods Local features Context feature Discourse feature Cross-document feature Neural network DMCNN
Motivation Patterns and features are equally important Although patterns cannot cover all representations of an event, it is still a very important feature. Candidate arguments can interact each other (1) Positive correlation: if one candidate argument belongs to one event, then the other is more likely to belong to the same event. (2) Negative correlation: if one candidate argument belongs to one event, then the other is less likely to belong to the same event.
Flow chart: (Grishman et al, 2005) Flow chart: Ours Flow chart: (Grishman et al, 2005)
Balancing the Pattern effects Feature set: pattern feature, trigger embedding and sentence-level embedding. Pattern feature : Each pattern has a corresponding event type. A candidate trigger may match more than one patterns, so that it has an event type distribution. Sentence-level embedding: extract all the NPs in the sentence and take the average word embedding of these NPs’ head word
Capturing two Relationships Between Arguments Positive correlation: if one candidate argument belongs to one event, then the other is more likely to belong to the same event; Negative correlation: if one candidate argument belongs to one event, then the other is less likely to belong to the same event. Relation matrix : C. If C(i,j)= 1, then argument i and argument j should belong to the same event. If C(i,j)=-1, then argument i and argument j cannot belong to the same event.
Evaluation function n-dim vector X: the identification result of arguments : the sum of all chosen arguments probability : the sum of all the classified roles’ probability
Training the Argument Relationship Matrix Features : TRIGGER: the trigger of the event ENTITY DISTANCE: the distance between the two candidate arguments in the sentence Whether the two candidate arguments occur on the same side of the trigger PARENT DEPENDENCY DISTANCE: the distance between the two candidate arguments’ parents in the dependency parse tree PARENT POS: if the two candidate arguments share the same parent, take the common parent’s POS tag as a feature Whether the two candidate arguments occur on the same side of the common parent if the two candidate arguments share the same parent MaxEnt classifier
Beam search algorithm We use the Beam Search algorithm to search for the assignment X:
Experiments Predicted entities, timex, values
Experiments Gold entities, timex, values
Thank you