Coreference Resolution in FreeLing 4.0

Coreference Resolution in FreeLing 4.0 Montserrat Marimon∗, Llu´ıs Padr´o†, Jordi Turmo† ∗Universitat Pompeu Fabra. Barcelona, Spain. montserrat.marimon@upf.edu †TALP Research Center. Universitat Polit`ecnica de Catalunya. Barcelona, Spain. {padro,turmo}@cs.upc.edu Abstract This paper presents the integration of RelaxCor into FreeLing. RelaxCor is a coreference resolution system based on constraint satisfaction that ranked second in the CoNLL-2011 shared task. FreeLing is an open-source library for NLP with more than fifteen years of existence and a widespread user community. We present the difficulties found in porting RelaxCor from a shared task scenario to a production enviroment, as well as the solutions devised. We present two strategies for this integration and a rough evaluation of the obtained results. Keywords: FreeLing, Coreference Resolution, CoNLL-2011, relaxation labeling 1. Introduction FreeLing1 is an open-source multilingual language process- ing library providing a wide range of analysis functionali- ties for several languages. The project is conceived as a library that can be called from a user application in need of analysis services. The software is open-source, distributed under an GNU Affero General Public License2, and dual-licensed to companies that em- bed it in their commercial products or online services. The open-source approach has been very fruitful during the fifteen years of life of the project (the first version was re- leased on 2003). The amount of accumulated downloads during this time is over 200,000. Contributions from the user community combined with the increasing availability of open source language resources has made it possible to extend the number of supported languages from three (En- glish, Spanish and Catalan) to fourteen (adding German, French, Italian, Portuguese, Russian, Norwegian, Asturian, Welsh, Galician, Croatian and Slovene). FreeLing offers a wide variety of language processing mod- ules, though not all modules are available for all languages. Most relevant modules are: Language identification, tok- enization and sentence splitting, lemmatization, date/time detection, numbers detection, multiword expressions detec- tion, physical magnitudes detection, named entity recogni- tion and classification, PoS tagging, SAMPA phonetic en- coding, word sense disambiguation, shallow parsing, con- stituency parsing, dependency parsing, semantic role la- belling, coreference resolution, semantic graph extraction, and document summarization. More details about the FreeLing project can be found in (Padr´o and Stanilovsky, 2012) and in the online documen- tation in the project website. One remarkable extension in version 4.0 was the inclu- sion of a Coreference Resolution module, based on Relax- Cor, the second-ranked system in CoNLL-2011 shared task 1http://nlp.lsi.upc.edu/freeling 2http://www.gnu.org/copyleft/agpl.html (Sapena et al., 2011). However, academic shared tasks have a very specific sce- nario, which does not necessarily match the real-world set- tings in which a system like FreeLing is required to operate. Thus, considerable efforts must be devoted to the integra- tion of a module successful in the laboratory, such as Re- laxCor, in a production pipeline. In this paper we describe how this coreference resolution module was integrated in FreeLing (for English and Span- ish), as well as the encountered obstacles and solutions de- vised. We also present an alternative configuration for Re- laxCor using hand-written constraints instead of …

Content truncated to 500 words