Relation extraction broadly speaking refers to the task of relating entities present in a document. This can take on many specific forms, such as labeling the relation between two given entities, or finding all entity pairs that satisfy a relation, or even multiway relation extraction (also called Record Extraction)
Relation Extraction has been studied in depth at MUC (MUC-6, MUC-7) and ACE (ACE2, ACE 2004, ACE 2005, ACE 2007, ACE 2008) series of conferences and evaluations. For biomedical literature, the BioCreAtIvE II tasks have also been useful.
The most common type of Relation Extraction is binary (i.e. for two entities), and can take on one of the following specific forms:
- Finding the type of relationship between given entities in text
Although in one sense this problem is easier than entity extraction because we only need to make one prediction instead of a vector of predictions, it is still considered harder because it requires a variety of syntactic and semantic features, both local as well as nonlocal. This problem is solved using one of the following types of methods:
1) Feature based methods These methods extract a flat set of features from the input and then invoke an off the shelf classifier like a decision tree or a SVM.
2) Kernel based methods These methods design special kernels to capture similarity between entities with complicated structures.
3) Rule based methods These methods create propositional and first order rules over structures around the two entities.
- For a given entity and relationship, find all entities that satisfy the relationship with the given entity.
- For a given relationship, find all entity pairs that satisfy it.
This problem is tackled by using a seed database of entity pairs to learn extraction patterns, which are then used to create candidate triplets (Entity1, Entity2, Relationsip). Finally, these candidates are pruned.
State of the art
In spite of the extensive research on the topic, the accuracy values of relation extraction systems in ACE evaluations still range only in the 50-70% neighborhood. Open-domain systems do even worse, and aren't usually based on very principled approaches; they contain a lot of special case-handling.
The Information Extraction Survey by Sunita Sarawagi contains more detail on this problem, and prior work to solve it.