The next step after identifying entities is to determine how they are related. There are different types of relations and each relation has their own subtypes. Relations take exactly two arguments, and the order of the arguments matters! For example:
George Bush travelled to France on Thursday for a summit.
There are two entities: “George Bush” and “France” and the relation between them is “Physical”. Within the relation “Physical”, it belongs to the subtype “Located”. This could be formulated as:
Relations are usually annotated between entity mentions as shown above, however, it can also hold between nominals. For example:
The cup contained tea from dried ginseng.
The relation above between tea and ginseng belongs to “Entity-Origin”. Nominal relation extraction is closely related to semantic role labelling (another NLP task). Relation extraction is usually restricted to small number of relation types restricted by the datasets.
There are multiple ways to extract relations. We will be exploring the following ways:
- Pattern-based relation extraction
- Relation extraction as a classification task
- Knowledge base population
- Open information extraction