Adapting an Entity Centric Model for Portuguese Coreference Resolution

Adapting an Entity Centric Model for Portuguese Coreference Resolution Evandro B. Fonseca1, Renata Vieira1, Aline Vanin2 1PUCRS University – Porto Alegre – Brazil 2UFCSPA University – Porto Alegre – Brazil evandro.fonseca@acad.pucrs.br, renata.vieira@pucrs.br, aline.vanin@ymail.com Abstract This paper presents the adaptation of an Entity Centric Model for Portuguese coreference resolution, considering 10 named entity categories. The model was evaluated on named e using the HAREM Portuguese corpus and the results are 81.0% of precision and 58.3% of recall overall, the resulting system is freely available. Keywords: Coreference Resolution, Information Extraction, Rule-based Models 1. Introduction Coreference resolution is a well known challenge in computational linguistics. While we could imagine that it is relatively easy to identify automatically that there is a link between identical or similar referents, such as in (1) “Barack Obama” and (2) “Obama”, this might not be always the case. See (3) “Adalberto Por- tugal” and (4) “Portugal”, for instance, whereas “Adal- berto Portugal” refers to a person, “Portugal” might re- fer either to “Adalberto” or to Portugal, the country. When dealing with Portuguese, this task is even more challenging, since resources are limited. Whereas Ontonotes (Pradhan et al., 2011), a coreference an- notated corpus for English, Chinese and Arabic has around 34.290 coreference chains, a corpus with sim- ilar purposes for Portuguese, Harem (Freitas et al., 2010), has approximately 887 coreference chains. In order to make available more resources for Por- tuguese, this paper presents the implementation and evaluation of an entity centric model for Portuguese, using a set of adapted rules, inspired by the Stanford Deterministic Coreference Resolution System (Lee et al., 2013). A rule based system was considered the best option, due to the shortage of annotated examples, nec- essary for learning approaches. We used the the Harem corpus (Freitas et al., 2010) to evaluate the adapted model. Our implementation is open source and we rely also on other open source resources, like Cogroo API (Silva, 2013). The rest of this paper is organized as follows: Section 2 presents the related work; Section 3 describes our version of the system for Portuguese; in Section 4 we present the evaluation of the system; in Section 5 con- clusions and future work are presented. 2. Related Work Coreference Resolution is an old topic in NLP, but it is still challenging. There are many studies involv- ing machine learning approaches. However, we con- sider that developing a rule based system would be more useful for Portuguese, since there is lack of a rich corpora, such as those employed to generate En- glish models (see (Fernandes et al., 2014), (Martschat and Strube, 2015), among others). Instead, we used an available Portuguese corpus for evaluating the adapted model. Since, we are mainly interested in rule-based approaches we focus here in describing the most influ- ential work for our system. (Lee et al., 2013) use a deterministic approach to coref- erence resolution that combines the global information and precise features previously identified by machine- learning models with the transparency and modular- ity of …

Content truncated to 500 words