Co-reference Resolution Module

The coreference system has been developed to decide wether two mentions refer to the same entity or not. The input of the systems consists of a list of Named Entities of type Person, i.e. Person Names (PNs), that have been automatically recognized in a document collection. Its output consists of a number of clusters of PNs, where each cluster is interpreted as the set of PNs that refer to the same entity.

The system has three main modules, depicted in Figure 1, corresponding to three steps of analysis: Person Name Splitter, Local Coreference module and Global Coreference module.

• The Person Name Splitter splits the Person Names into first_name and last_name. The motivation is that first_name and last_name are used differently in mentioning persons and play different roles in coreference. We have implemented a simple Multilayer Perceptron Network (see fig.2).
• The Local Coreference module tries to establish among the PNs occurring within a single document which ones refer. We use a probabilistic approach, estimating, inside each document, the joint probability of occurrence of any pair formed by two free names (either first or last names) and a certain complete name (first and last name). For each cluster of coreferred PNs, we choose a representing name, which we call local head. The set of local heads is the output of this module. The probabilities are computed according to the formula presented in figure 3.
• The Global Coreference module tries to establish the coreference among all the local heads of the collection based on their association set (the set of named entities they are associated with) and on the probability that the same PN refers to two different persons (common vs. uncommon names).

max_L p(l , L , l_L , t) = max_L p(t) p(L | t) p(l | L, t) p(l_L | l , L , t)

Fig. 3: Probability formula, where l stands for a first name, L for last name, l_L for first last name combination and t for topic.

• The Name Dictionary is a list of first_names and last_names. We start with a Name Dictionary of Italian names which contains 68,654 last_names and 3,230 first_names. There are 1267 ambiguous names, which can be both first and last name.
• The Entity Ontology contains all the known entities. We start with an empty Entity Ontology, which, at the end of the coreference process, will contain all the person entities identified in the corpus.
• The Topic Ontology is a resource specific for the news domain, where each document of the collection is classified under a limited number of topics. The idea is that PNs that occurs in the same topic tend to increase their probability to corefer. We start with a fix set of topics, which are actually the newspaper sections.
  

  
  
  Enter the Ontotext Portal
  (access is restricted to the partners of the project)

  
  
  
  Publications
  
  

  ---

  Maintainer: bentivofbk.eu
  Last modified: Tue Aug 28 2007