In event detection, a schema is provided for each type of event (e.g., an election, terrorist attack, etc), indicating all the possible properties of the event. The system is required to fill in as many of these properties as possible.
Event detection systems generally have two components:
Retrieval – Identify relevant documents and passages of text
Extraction – Identify the properties of the event based on the retrieved texts
Early approaches focused on finite-state patterns for identify event properties. Feature engineered patterns that are most likely to appear in documents when a specific event query is made.
Other approaches include using bidirectional RNN or techniques that are used in semantic role labelling such as FrameNet. Both can be used for structured prediction over local and global features.
It is possible that multiple sentences describe properties of a single event. Therefore, event coreference is needed to connect all the event mentions. Event coreference can be define as the task of identifying event mentions that share the same event participants (entities) and properties.
Supervised learning can be used for event coreference resolution, where you can move left-to-right through the document and use a classifier to determine whether to link each event to an existing cluster of coreferent events, or to create a new cluster. Each clustering is based on the entities and properties of the event.
Due to lack of annotated data, unsupervised approaches are more desirable.
Relation between events
Events may be related to other events, both temporally and/or causally. The TimeML annotation scheme has a set of six temporal relations between events. The TimeBank corpus has 186 annotated documents. With these annotated data, we can use supervised learning with temporal constraints to detect temporal relations.