Sanaphor++: Combining Deep Neural Networks with Semantics for Coreference Resolution

Sanaphor++: Combining Deep Neural Networks with Semantics for Coreference Resolution Julien Plu♦, Roman Prokofyev∗, Alberto Tonon∗, Philippe Cudr´e-Mauroux∗, Djellel Eddine Difallah∗, Rapha¨el Troncy♦, Giuseppe Rizzo∞ ♦EURECOM, France {julien.plu,raphael.troncy}@eurecom.fr ∗eXascale Infolab, University of Fribourg {roman.prokofyev,alberto.tonon,philippe.cudre-mauroux,djelleleddine.difallah}@unifr.ch ∞ISMB, Italy giuseppe.rizzo@ismb.it Abstract Coreference resolution has always been a challenging task in Natural Language Processing. Machine learning and semantic techniques have improved the state of the art over the time, though since a few years, the biggest step forward has been made using deep neural networks. In this paper, we describe Sanaphor++, which is an improvement of a top-level deep neural network system for coreference resolution—namely Stanford deep-coref—through the addition of semantic features. The goal of Sanaphor++ is to improve the clustering part of the coreference resolution in order to know if two clusters have to be merged or not once the pairs of mentions have been identified. We evaluate our model over the CoNLL 2012 Shared Task dataset and compare it with the state-of-the-art system (Stanford deep-coref) where we demonstrated an average gain of 1.13% of the average F1 score. Keywords: Deep Learning, Coreference Resolution, Entity Linking 1. Introduction The task of coreference resolution aims to identify which mentions in a text refer to the same real-world entity. Al- though coreference resolution is mostly studied as a clus- tering problem, it has also been studied as a Semantic Web problem by using Named Entity Recognition (NER) and Named Entity Linking (NEL) approaches. We define a Se- mantic Web problem as a problem where we exploit the se- mantics represented in a knowledge base that is published on the Web. Coreference resolution is an important as- pect of text understanding and has numerous applications such as Entity Linking (see, for instance, the first two edi- tions of the Open Knowledge Extraction challenge (Nuz- zolese et al., 2015; Nuzzolese et al., 2016)). As an ex- ample of coreference resolution, in the following piece of text “Emmanuel Macron is the new French president. He has been elected with a large majority”, the mentions Emmanuel Macron and He will be disambiguated to the same entity, e.g. http://dbpedia.org/resource/ Emmanuel_Macron. The task of coreference resolution is considered as one of the most challenging in Natural Lan- guage Processing (NLP). As an example of its challenging nature, in the following sentence “Curie shared the 1903 Nobel Prize in Physics with her husband, Pierre Curie”, it is clear that the clusters {Pierre Curie} and {Curie, her} are disjoint and do not refer to the same entity, but it is ambiguous whether the pair of mentions Pierre Curie and Curie are coreferent or not. Actually, without the mention her, one does not know if the mention Curie refers to Pierre Curie or Marie Curie. The contributions of this work are: 1. A new approach that leverages both deep learning and Semantic Web techniques to solve an NLP problem; 2. A model that is integrated into a widely used NLP toolkit, namely the Stanford CoreNLP; 3. A thorough evaluation showing that our technique improves …

Content truncated to 500 words