What is Coreference Resolution?
Coreference Resolution is the task of finding all the mentions that refer to the same entity in a source document.
What is Anaphoric and Cataphoric references?
Anaphoric references occurs when a word refers back to an entity that has already been introduced. For example, Ryan breaks the glass by knocking it off the table. The word ‘it’ is referring back to the glass entity that has already been introduced.
Cataphoric references occurs when a word refers to an entity that’s being introduce at a later stafe. For example, When I visit her, my mum cooks me delicious food. ‘Her’ refers to the entity mum that’s only introduce at the later stage of the text.
Why is this important?
The Winogard Schema Challenge is the improvement of the Turing test designed to measure machine intelligence. Winograd Schema questions require machine to be able to perform anaphoric resolutions. This makes coreference resolution the basis of the Winograd Schema Challenge.
How does coreference resolution algorithm works in general?
Extract all the mentions that are referring to entities
For each mention and pair of mentions, we compute a set of features
Identify the most likely antecedent for each mention using the set of features. This is also known as pairwise ranking
Describe a typical neural network for coreference scoring.
Average the word embeddinggs of words inside and around each mention
Add simple integer features to compute features representation for each mention and its surroundings
We feed these representations into two neural networks. The first network score each pair of a mention and a possible antecedent and the second network scores each mention having no antecedent
Compare these scores and take the highest score to determine if a mention has an antecedent or not and if it does, which antecedent is the most appropriate