Journal of Artificial Intelligence Research 40 (2011) 469–521 Submitted 06/10; published 02/11
Altaf Rahman altaf@hlt.utdallas.edu
Vincent Ng vince@hlt.utdallas.edu
Human Language Technology Research Institute
University of Texas at Dallas
800 West Campbell Road; Mail Station EC31
Richardson, TX 75080-3021 U.S.A.
http://www.jair.org/media/3120/live-3120-5478-jair.pdf
-Long paper, very thorough
-A lot of history, use the citations in this paper
-Learn more about "centering algorithms"
-Describes three different models, mention-pair,entity-mention and mention-ranking
-Outlines key features and deficiencies of each
-In particular the transitivity property is not addressed in the mention-pair model so clustering is used
-Mention-ranking outperforms mention-pair
-Describe a cluster-ranking approach combines both models
-Use lexicalization and knowledge of anaphoricity
-Used ACE for experiments
Interesting:
"Specifically, a classifier that is trained on
coreference-annotated data is used to determine whether a pair of mentions is co-referring
or not. However, the pairwise classifications produced by this classifier (which is now commonly
known as the mention-pair model) may not satisfy the transitivity property inherent
in the coreference relation, since it is possible for the model to classify (A,B) as coreferent,
(B,C) as coreferent, and (A,C) as not coreferent. As a result, a separate clustering mechanism
is needed to coordinate the possibly contradictory pairwise classification decisions and
construct a partition of the given mentions."
Read about Lappin and Leass’s algorithm
Read about centering algorithms
"the distinction between
classification and ranking applies to discriminative models but not generative models.
Generative models try to capture the true conditional probability of some event. In the context
of coreference resolution, this will be the probability of a mention having a particular
antecedent or of it referring to a particular entity (i.e., preceding cluster). Since these probabilities
have to normalize, this is similar to a ranking objective: the system is trying to raise
the probability that a mention refers to the correct antecedent or entity at the expense of
the probabilities that it refers to any other. Thus, the antecedent version of the generative
coreference model as proposed by Ge et al. (1998) resembles the mention-ranking model,
while the entity version as proposed by Haghighi and Klein (2010) is similar in spirit to the
cluster-ranking model."
