Annotating Abstract Meaning Representations for Spanish

Annotating Abstract Meaning Representations for Spanish Noelia Migueles-Abraira, Rodrigo Agerri, Arantza Diaz de Ilarraza IXA NLP Group, University of the Basque Country UPV/EHU noeliamigueles@gmail.com,{rodrigo.agerri,a.diazdeilarraza}@ehu.eus Abstract Until recently, semantic annotations for different semantic phenomena were independent and unconnected. The Abstract Meaning Representation (AMR) project arised out of the need to create a broad-coverage semantic bank containing a unified set of semantic information represented in simple, single-rooted, easy-to-read structures. Because the semantic representation language proposed in AMR is biased towards English, annotating AMR structures for other languages, such as Spanish, is not a trivial task. In this paper we propose a linguistic method that we believe would help lay the groundwork for building a large semantic bank for Spanish and would guide those who would like to implement it for other languages. Thus, we analyze a broad spectrum of Spanish linguistic phenomena to come up with suggestions to adapt the current guidelines so that it is possible to annotate AMRs for Spanish. As a result of this work, we make available the first public online repository containing manually annotated Spanish AMRs. Keywords: AMR, Sembank, Semantics, Spanish Language Resources 1. Introduction The construction of large manually annotated corpora has been crucial in order to advance Natural Language Pro- cessing (NLP) technology. Among others, the construction of the Penn Treebank (Marcus et al., 1993) made possible the existence of many statistical parsers trained on syntac- tic treebanks which perform at an accuracy of about 90% (Charniak and Johnson, 2005). In contrast, computational semantics is lagging behind in this respect, as most of the semantic annotated resources for English are rather scat- tered or simply nonexistent for the large majority of lan- guages. This lack of a unified sembank of natural language sen- tences paired with their sentential, logical meanings is what led to the appearance of Abstract Meaning Representa- tion (AMR), a semantic representation language introduced by (Banarescu et al., 2013). This approach promotes the representation of the logical meaning of sentences as sin- gle rooted, directed graphs (or AMRs) with labelled nodes (concepts) and edges between them (relations). These in- corporate semantic roles, among other linguistic phenom- ena. In a propositional-style logic, AMR is able to cap- ture who is doing what to whom in a sentence. Further- more, AMR tries to abstract away from both morphological and syntactic idiosyncrasies that account for several cross- lingual differences. However, there is still a problem with this approach, as it is strongly biased (by design) to anno- tate English sentences. To overcome this fundamental lim- itation, we decided to explore the possibility of annotating AMRs for Spanish texts. Thus, the central goal of this work is to investigate how to create Spanish AMRs in order to build a sizable Spanish sembank. We address this objective by answering the fol- lowing questions: 1. Is it possible to follow the current guidelines to anno- tate Spanish AMRs? And if not, how can the guide- lines be refined in order to annotate Spanish AMRs? Also, what resources do …

Content truncated to 500 words