Deep Neural Networks for Coreference Resolution for Polish

Deep Neural Networks for Coreference Resolution for Polish Bartłomiej Nito´n, Paweł Morawiecki, and Maciej Ogrodniczuk Institute of Computer Science, Polish Academy of Sciences, Warsaw, Poland {bartek.niton, pawel.morawiecki}@gmail.com, maciej.ogrodniczuk@ipipan.waw.pl Abstract The paper presents several configurations of deep neural networks aimed at the task of coreference resolution for Polish. Starting with the basic feature set and standard word embedding vector size we examine the setting with larger vectors, more extensive sets of mention features, increased number of negative examples, Siamese network architecture and a global mention clustering algorithm. The highest results are achieved by the system combining our best deep neural architecture with the sieve-based approach – the cascade of rule-based coreference resolvers ordered from most to least precise. All systems are evaluated on the data of the Polish Coreference Corpus featuring 540K tokens and 180K mentions. The best variant improves the state of the art for Polish by 0.53 F1 points, reaching 81.23 points of the CoNLL metric. Keywords: coreference resolution, deep neural network, Polish 1. Introduction Coreference resolution, the task of clustering textual frag- ments that refer to the same entity in the discourse world, has been successfully tackled for Polish in numerous configurations, starting with a rule-based model (Ogrod- niczuk and Kope´c, 2011) through machine-learning (Kope´c and Ogrodniczuk, 2012) and projection-based approaches (Ogrodniczuk, 2013) up to the newest multi-pass sieve set- ting (Nito´n and Ogrodniczuk, 2017). In this paper we present the first deep neural network resolver for Polish, further improving the state of the art. For English, the state- of-the-art coreference resolution systems are also based on deep neural networks (Clark and Manning, 2016), (Wise- man et al., 2016). We were inspired and motivated by these works. The data for our experiments, as for all previous configu- rations, come from the Polish Coreference Corpus (Ogrod- niczuk, 2015, PCC), a large corpus of Polish general nom- inal coreference manually annotated over the texts of the National Corpus of Polish (Przepiórkowski et al., 2012)1 and Rzeczpospolita Corpus (Presspublica, 2002). The cor- pus features broad understanding of mentions (e.g. with included relative clauses or appositions, nesting, disconti- nuities and zero anaphora) and contains almost 1800 doc- uments from 14 genres, 540K tokens, 180K mentions and 128K coreference clusters. Coreference scores on the test set are measured using gold mentions on input with MUC (Vilain et al., 1995), B3 (Bagga and Baldwin, 1998), and CEAFE (Luo, 2005) metrics averaging them according to the CoNLL-2011 ap- proach (Pradhan et al., 2011) to track influence on differ- ent coreference dimensions (the B3 measure being based on mentions, MUC on links, and CEAFE on entities). CEAFM (Luo, 2005) and BLANC (Recasens and Hovy, 2011) are also presented for consideration. Metrics were calculated using Scoreference2, a mention detection and coreference resolution evaluation tool (Ogrodniczuk et al., 2015). 1Pol. Narodowy Korpus J˛ezyka Polskiego (NKJP), see http: //nkjp.pl. 2http://zil.ipipan.waw.pl/Scoreference 2. The Baseline In all our experiments we used 90% texts from the PCC as the training set and 10% as the test set. Text type balance was maintained …

Content truncated to 500 words