On The Feasibility of Open Domain Referring Expression Generation Using Large Scale Folksonomies

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 641–645, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics On The Feasibility of Open Domain Referring Expression Generation Using Large Scale Folksonomies Fabi´an Pacheco Pablo Ariel Duboue∗ Facultad de Matem´atica, Astronom´ıa y F´ısica Universidad Nacional de C´ordoba C´ordoba, Argentina Mart´ın Ariel Dom´ınguez Abstract Generating referring expressions has received considerable attention in Natural Language Generation. In recent years we start seeing deployments of referring expression genera- tors moving away from limited domains with custom-made ontologies. In this work, we ex- plore the feasibility of using large scale noisy ontologies (folksonomies) for open domain referring expression generation, an important task for summarization by re-generation. Our experiments on a fully annotated anaphora resolution training set and a larger, volunteer- submitted news corpus show that existing al- gorithms are efficient enough to deal with large scale ontologies but need to be extended to deal with undefined values and some mea- sure for information salience. 1 Introduction Given an entity1 (the referent) and a set of com- peting entities (the set of distractors), the task of referring expression generation (REG) involves cre- ating a mention to the referent so that, in the eyes of the reader, it is clearly distinguishable from any other entity in the set of distractors. In a traditional generation pipeline, referring expression generation happens at the sentence planning level. As a result, its output is not a textual nugget but a description employed later on by the surface realizer. In this pa- per, we consider the output of the REG system to ∗To whom correspondence should be addressed. Email: pablo.duboue@gmail.com. 1Or set of entities, but not in this work. be Definite Descriptions (DD) consisting of a set of positive triples and a set of negative triples, enumer- ating referent-related properties. Since the seminal work by Dale and Re- iter (1995), REG has received a lot of attention in the Natural Language Generation (NLG) community. However, most of the early work on REG has been on traditional NLG systems, using custom-tailored ontologies. In recent years (Belz et al., 2010) there has been a shift towards what we term “Open Do- main Referring Expression Generation,” (OD REG), that is, a REG task where the properties come from a folksonomy, a large-scale volunteer-built ontology. In particular, we are interested in changing anaphoric references for entities appearing in sen- tences drafted from different documents, as done in multi-document summarization (Advaith et al., 2011). For example, consider the following sum- mary excerpt2 as produced by Newsblaster (McKe- own et al., 2002): Thousands of cheering, flag-waving Palestinians gave Palestinian Authority President Mahmoud Abbas an en- thusiastic welcome in Ramallah on Sunday, as he told them triumphantly that a “Palestinian spring” had been born following his speech to the United Nations last week.3 The president pressed Israel, in unusually frank terms, to reach a final peace agreement with the Pales- tinians, citing the boundaries in place on the eve …

Content truncated to 500 words