“One Entity per Discourse” and “One Entity per Collocation” Improve Named-Entity Disambiguation

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2260–2269, Dublin, Ireland, August 23-29 2014. “One Entity per Discourse” and “One Entity per Collocation” Improve Named-Entity Disambiguation Ander Barrena*, Eneko Agirre*, Bernardo Cabaleiro**, Anselmo Pe˜nas**, Aitor Soroa* *IXA NLP Group / University of the Basque Country, Basque Country abarrena014@ikasle.ehu.es, e.agirre@ehu.es, a.soroa@ehu.es **UNED NLP & IR Group, Madrid anselmo@lsi.uned.es, bcabaleiro@lsi.uned.es Abstract The “one sense per discourse” (OSPD) and “one sense per collocation” (OSPC) hypotheses have been very influential in Word Sense Disambiguation. The goal of this paper is twofold: (i) to explore whether these hypotheses hold for entities, that is, whether several mentions in the same discourse (or the same collocation) tend to refer to the same entity or not, and (ii) test their impact in Named-Entity Disambiguation (NED). Our experiments show consistent results on different collections and three state-of-the-art NED system. OSPD hypothesis holds in around 96%-98% of documents whereas OSPC hypothesis holds in 91%-98% of collocations. Furthermore, a simple NED post-processing in which the majority entity is promoted, produces a gain in perfor- mance in all cases, reaching up to 8 absolute points of improvement in F-measure. These results show that NED systems would benefit of considering these hypotheses into their implementation. 1 Introduction The “one sense per discourse” (OSPD) hypothesis was introduced by Gale et al. (1992), and stated that a word tends to preserve its meaning when occurring multiple times in a discourse. They estimated that the probability of two occurrences of the same polysemous noun drawn from one document having the same sense to be around 94% for documents from Grolier encyclopedia, and 96% for documents from Brown, based on word senses from the Oxford Advanced Learner’s Dictionary and a handful of examples. A few years later, Krovetz (1998) reported 66% on larger corpora (SemCor and DSO) annotated with WordNet senses by third parties, but, unfortunately, he only reported how many polysemous nouns occurred with a single sense in all documents, not in each document. In the context of statistical machine translation, Carpuat (2009) reported that, 80% of the time, words occurring multiple times in a source document are translated into a single word in the target language. In the case of entities, OSPD is closely related to coreference, where the task is to find whether two different mentions (perhaps using different surface strings like “John” and “he”) in a document refer to the same entity or not. For instance, the coreference system presented by (Lee et al., 2013), uses a heuristic which links mentions in a document that share the same surface string: “This sieve [heuristic] accounts for approximately 16 CoNLL F1 points improvement, which proves that a signicant percentage of mentions in text are indeed repetitions of previously seen concepts”. Our paper actually quantifies the amount of those repetitions for entities, providing additional evidence for the heuristic. The “one sense per collocation” (OSPC) hypothesis was introduced by Yarowsky (1993), stating that a word tends to preserve its meaning when …

Content truncated to 500 words