Spanish Abstract Meaning Representation: Annotation of a General Corpus

Spanish Abstract Meaning Representation: Annotation of a General Corpus Shira Wein, Georgetown University, Washington, DC, USA sw1158@georgetown.edu Lucia Donatelli, Saarland University, Germany donatelli@coli.uni-saarland.de Ethan Ricker, Georgetown University, Washington, DC, USA ear131@georgetown.edu Calvin Engstrom, Georgetown University, Washington, DC, USA cle41@georgetown.edu Alex Nelson, Georgetown University, Washington, DC, USA amn106@georgetown.edu Leonie Harter, Saarland University, Germany leonie-harter@web.de Nathan Schneider, Georgetown University, Washington, DC, USA nathan.schneider@georgetown.edu Abstract Abstract Meaning Representation (AMR), originally designed for English, has been adapted to a number of languages to facilitate cross-lingual semantic representation and analysis. We build on previous work and present the first sizable, general annotation project for Spanish AMR. We release a detailed set of annotation guidelines and a corpus of 486 gold-annotated sentences spanning multiple genres from an existing, cross-lingual AMR corpus. Our work constitutes the second largest non-English gold AMR corpus to date. Fine-tuning an AMR-to-Spanish generation model with our annotations results in an absolute BERTScore improvement of 8.8%, demonstrating initial utility of our work. 1 Introduction Abstract Meaning Representation (AMR) represents the core meaning of a sentence as a directed, rooted graph focused on predicate-argument structure (Ba- narescu et al., 2013) (figure 1). Nodes correspond to concepts and labels denote relations between concepts. Labels can be core roles functioning as predicates or arguments, or other attributes such as :location or :manner. While there are large AMR-annotated corpora available for English, cross-lingual adaptations of AMR are necessary if AMR is to be useful as an in- terlingua or intermediate representation for cross- lingual tasks (Xue et al., 2014). Recent work has adapted AMR to a variety of languages (§2.1), evalu- ating cross-lingual efficacy of rolesets, word senses, and how effectively AMR relations capture “who is doing what to whom” in languages other than English. As AMR aims to abstract away from morphosyn- tax, its graph structure is closer to logic than a syn- tactic parse. For English, AMR removes information such as number, definiteness, tense, word class, and word order. Yet, in many languages, morphosyn- tactic information in languages other than English carries rich, important semantic information be- yond the “sugar” AMR intends to avoid. Therefore, it is important when developing non-English AMR annotation schema to both consider consistency with work in other languages (primarily English) as well as effectively reflecting the semantics of the language being annotated as much as possible. Spanish is one of the most widely spoken lan- guages in the world. There has been one previous proposal for adapting AMR to Spanish: Migueles- Abraira et al. (2018) presented a corpus of 50 rep- resentative annotations for a Spanish translation of (The Little Prince) (LPP) (§2.2). While Migueles- Abraira et al. (2018) noted that English AMR failed Northern European Journal of Language Technology (a) (s / say-01 :ARG0 (p / prince :mod (l / little)) :ARG1 (h / hurry-01) :ARG1 (t / they) :degree (g / great)) (b) (d / decir-01 :ARG0 (p / príncipe :mod (p2 / pequeño)) :ARG1 (a / apresurado :domain (t / th-pers-pl-sinnombre) :degree (m / muy))) Figure 1: …

Content truncated to 500 words