Coreference Resolution and Entity Linking

Coreference Resolution and Entity Linking

Pairwise, ranking and span-based resolution

Scoring chains with MUC and B3

Linking names to the world

In It rained hard during the crossing, and the skipper logged it in the book, what should a mention detector do with the two occurrences of it?

In the end-to-end span-ranking coreference model, the antecedent variable ranges over earlier spans plus a dummy $\epsilon$ whose score is fixed at 0. What does scoring $\epsilon$ at exactly 0 accomplish?

An end-to-end coreference model considers every contiguous span in a $21$-word document before length truncation. How many candidate spans are there?

For a given mention, the hypothesis chain containing it has $2$ mentions, of which $1$ also appear in the gold chain, and the gold chain has $5$ mentions. Compute the mention's B3 recall, rounding to three decimals.

A system splits one gold four-mention chain into two two-mention chains and leaves a gold singleton untouched. How do MUC and B3 respond?