Abstract Meaning Representation of Constructions: The More We Include, the Better the Representation

Abstract Meaning Representation of Constructions: The More We Include, the Better the Representation Claire Bonial,1 Bianca Badarau,2 Kira Griffitt,3 Ulf Hermjakob,4 Kevin Knight,4 Tim O’Gorman,5 Martha Palmer,5 Nathan Schneider6 1U.S. Army Research Laboratory, 2SDL, 3Linguistic Data Consortium, 4USC Information Sciences Institute, 5University of Colorado, 6Georgetown University claire.n.bonial.civ@mail.mil Abstract We describe the expansion of the Abstract Meaning Representation (AMR) project to provide coverage for the annotation of certain types of constructions. Past AMR annotations generally followed a practice of assigning the semantic roles associated with an individual lexical item, as opposed to a flexible pattern or template of multiple lexical items, which characterizes constructions such as ‘The X-er, The Y-er’ (exemplified in the title). Furthermore, a goal of AMR is to provide consistent semantic representation despite language-specific syntactic idiosyncracies. Thus, representing the meanings associated with fully syntactic patterns required a novel annotation approach. As one strategy in our approach, we expanded the AMR lexicon of predicate senses, or semantic ‘rolesets,’ to include entries for a growing set of constructions. Despite the challenging practical and theoretical questions encountered, the additions and updates to AMR annotation described here ensure more comprehensive semantic representations capturing both lexical and constructional meaning. Keywords: Semantics, Constructions, Meaning Representation 1. Introduction The Abstract Meaning Representation (AMR) project (Ba- narescu et al., 2013) has created a manually annotated “se- mantics bank” of text. A goal of AMR is to capture core facets of meaning while abstracting away from idiosyncratic syntactic facts; thus, for example, She adjusted the machine and She made an adjustment to the machine share the same AMR. The purpose of these annotations is to support natural language processing (NLP) applications such as natural lan- guage understanding, generation, and summarization (Liu et al., 2015; Pourdamghani et al., 2016), machine translation, question answering (Mitra and Baral, 2016), information extraction (Pan et al., 2015), and biomedical text mining (Garg et al., 2016; Rao et al., 2017; Wang et al., 2017). With a growing body of over 70 research papers using AMR,1 the corpus is becoming a benchmark dataset. As a practical NLP resource, AMR annotation has focused on providing coverage for the more frequent and predictably patterned linguistic phenomena, and therefore has not nec- essarily provided adequate representations for some of the rarer structures found in the long tail of language (Zipf, 1949). However, as the project matures, we are aiming to ex- pand the representation to go beyond capturing the semantics of purely compositional language to better capture the se- mantics of constructions in which the meaning is more than the sum of its parts (lexical meanings). From its inception, in ‘abstracting away’ from language-specific syntactic facts to represent core meaning elements, AMR has annotated a variety of semi- and non-compositional phrases accord- ing to a more general lexicon entry capturing the meaning, independent of any particular syntactic derivation. For ex- ample, certain realizations of but, while, whereas, however 1Bibliographies: https://amr.isi.edu/research.html, http://people.cs.georgetown.edu/nschneid/cosc672/s17/ amr-papers.html and on the other hand are represented using a single lexi- con entry, contrast-01. Nonetheless, …

Content truncated to 500 words