Inferring a Noun Phrase Clustering for an Antecedent-based Co-reference Resolution Model
Jason D. M. Rennie · 2004
We assume that a set of noun phrases {x1, . . . , xn} have been extracted from a text. We also assume the parameters for our model, ~ w, have been learned (Rennie, 2004). We would like to determine a configuration (set of labels), ~y, that maximizes the joint likelihood of the model. Since our model is probabilistic, approximate inference can be achieved via belief propagation (BP). However, since the underlying graph is fully connected, BP is not very useful. We instead consider a set of simpler, greedy algorithms. The first algorithm we consider is also the simplest. We call it MaxAntecedent. Noun phrases are ordered according to their appearance in text. In order, a label is chosen for each noun phrase according to the maximum likelihood antecedent. That is, each noun phrase takes on the label of the noun phrase that has the highest antecedent probability. The next algorithm is a variation on MaxAntecedent that is more in-line with our joint likelihood objective. We call it GreedyTopDown. Again, noun phrases are ordered according to their appearance in text. In order, labels are chosen to maximize conditional likelihood of the noun phrase label given the labels of all preceeding noun phrases. Note that if a noun phrase has many low-probability