Person Cross Document Coreference with Name Perplexity Estimates

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 997–1006, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Person Cross Document Coreference with Name Perplexity Estimates Octavian Popescu popescu@racai.ro . Abstract The Person Cross Document Coreference sys- tems depend on the context for making deci- sions on the possible coreferences between person name mentions. The amount of context required is a parameter that varies from cor- pora to corpora, which makes it difficult for usual disambiguation methods. In this paper we show that the amount of context required can be dynamically controlled on the basis of the prior probabilities of coreference and we present a new statistical model for the compu- tation of these probabilities. The experiment we carried on a news corpus proves that the prior probabilities of coreference are an impor- tant factor for maintaining a good balance be- tween precision and recall for cross document coreference systems. 1 Introduction The Person Cross Document Coreference (Grishman 1994) task, which requires that all and only the textual mentions of an entity of type Person be individuated in a large collection of text documents, is one of the challenging tasks for natural language processing systems. In the most general case the corpus itself is the only available source of information regarding the persons mentioned and we consider that this is the case in this paper. A PCDC system must be able to use the information existing in the corpus in order to assign to each personal name mention (PNM) a piece of context. The coreference of any two PNMs is decided mainly on the basis of the similarity of the pieces of contexts associated with them. A successful PCDC must accurately extract the relevant context for coreference. However, the context relevance is not abso- lute. Whether the contextual information uniquely individuates a person is a matter of probability. This paper presents a statistical tech- nique developed to provide a PCDC system with more information regarding the probability of a correct coreference. The reason for developing this technique is twofold: (i) the relevant corefer- ence context depends on the corpus itself and (ii) valid coreferences require a large amount of in- formation, which is unavailable in the majority of cases. The first reason is linked to a particularity of the CDC task that makes it more complex than other NLP tasks. Unlike in other disambiguation tasks, in the CDC tasks the relevant coreference context depends on the corpus itself. In word sense disambiguation, for instance, the distribu- tion of the relevant context is mainly regulated by strong syntactic and semantic rules. The exis- tence of such rules makes it possible for the dis- ambiguation decisions to be made considering the local context. On the other hand, the distribu- tion of the PNMs in a corpus is rather random and the relevant coreference context is a dynamic variable depending on the diversity of the corpus, that is, on how many different persons with the same name share a …

Content truncated to 500 words