Generative Event Schema Induction with Entity Disambiguation

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 188–197, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Generative Event Schema Induction with Entity Disambiguation Kiem-Hieu Nguyen1, 2 Xavier Tannier3, 1 Olivier Ferret2 Romaric Besanc¸on2 (1) LIMSI-CNRS (2) CEA, LIST, Laboratoire Vision et Ingnierie des Contenus, F-91191, Gif-sur-Yvette (3) Univ. Paris-Sud {nguyen,xtannier}@limsi.fr, {olivier.ferret,romaric.besancon}@cea.fr Abstract This paper presents a generative model to event schema induction. Previous meth- ods in the literature only use head words to represent entities. However, elements other than head words contain useful in- formation. For instance, an armed man is more discriminative than man. Our model takes into account this information and precisely represents it using proba- bilistic topic distributions. We illustrate that such information plays an important role in parameter estimation. Mostly, it makes topic distributions more coherent and more discriminative. Experimental results on benchmark dataset empirically confirm this enhancement. 1 Introduction Information Extraction was initially defined (and is still defined) by the MUC evaluations (Grish- man and Sundheim, 1996) and more specifically by the task of template filling. The objective of this task is to assign event roles to individual tex- tual mentions. A template defines a specific type of events (e.g. earthquakes), associated with se- mantic roles (or slots) hold by entities (for earth- quakes, their location, date, magnitude and the damages they caused (Jean-Louis et al., 2011)). Schema induction is the task of learning these templates with no supervision from unlabeled text. We focus here on event schema induction and con- tinue the trend of generative models proposed ear- lier for this task. The idea is to group together entities corresponding to the same role in an event template based on the similarity of the relations that these entities hold with predicates. For ex- ample, in a corpus about terrorist attacks, enti- ties that are objects of verbs to kill, to attack can be grouped together and characterized by a role named VICTIM. The output of this identification operation is a set of clusters of which members are both words and relations, associated with their probability (see an example later in Figure 4). These clusters are not labeled but each of them represents an event slot. Our approach here is to improve this initial idea by entity disambiguation. Some ambiguous enti- ties, such as man or soldier, can match two differ- ent slots (victim or perpetrator). An entity such as terrorist can be mixed up with victims when arti- cles relate that a terrorist has been killed by police (and thus is object of to kill). Our hypothesis is that the immediate context of entities is helpful for disambiguating them. For example, the fact that man is associated with armed, dangerous, heroic or innocent can lead to a better attribution and def- inition of roles. We then introduce relations be- tween entities and their attributes in the model by means of syntactic relations. The document level, which …

Content truncated to 500 words