AMR Beyond the Sentence: the Multi-sentence AMR corpus

Proceedings of the 27th International Conference on Computational Linguistics, pages 3693–3702 Santa Fe, New Mexico, USA, August 20-26, 2018. 3693 AMR Beyond the Sentence: the Multi-sentence AMR corpus Tim O’Gorman1, Michael Regan3, Kira Griffitt2, Ulf Hermjakob4, Kevin Knight4, Martha Palmer1 1University of Colorado Boulder {ogormant,mpalmer}@colorado.edu 2Linguistic Data Consortium kiragrif@ldc.upenn.edu 3University of New Mexico reganman@unm.edu 4Information Sciences Institute, University of Southern California {ulf,knight}@isi.edu Abstract There are few corpora that endeavor to represent the semantic content of entire documents. We present a corpus that accomplishes one way of capturing document level semantics, by annotat- ing coreference, implicit roles and bridging relations on top of gold Abstract Meaning Repre- sentations of sentence-level semantics. We present the methodology of developing this corpus, alongside analysis of its quality and a plausible baseline for comparison. It is hoped that this Multi-Sentence AMR corpus (MS-AMR) illustrates a feasible approach to developing rich rep- resentations of document meaning, useful for tasks such as information extraction and question answering. 1 Introduction Although Abstract Meaning Representation (AMR) (Banarescu et al., 2013) shows promise for a range of tasks such as summarization (Liu et al., 2015; Viet et al., 2017) and information extraction (Garg et al., 2016), it is restricted to capturing the semantics of individual sentences. For many purposes, when examining the semantics of a document, one also needs access to cross-sentence information such as coreference. We suggest that the AMR approach to semantic representation has useful characteristics for an exten- sion to discourse-level representations. AMR represents sentence meaning in a simple, readable semantic graph, such that annotators may directly mark coreference relations upon the AMR graph itself. AMR also annotates implicit roles in some within-sentence contexts, and the PropBank predicate annotations provide a resource for extending those implicit roles annotations to a document level. The Multi-Sentence Abstract Meaning Representation (MS-AMR) corpus is a corpus annotated on top of existing gold AMRs, extending them with this additional information. By linking those AMRs together, it presents an integrated representation of the meaning of an entire document or discourse, as the addition of the coreference, implicit role reference and bridging relations across each AMR helps to build a larger representation of the entire propositional content of the document. Because these MS- AMR representations are annotated directly onto the variables within an AMR semantic representation, it is also a different task from traditional coreference, event coreference or implicit role coreference tasks, and results in a fundamentally different kind of data. We present a baseline system and inter- annotator agreement scores, which we hope will illuminate the nature and quality of the dataset, and outline methods for how to score MS-AMR system outputs. 2 Annotation Methodology 2.1 Background: Within-sentence Abstract Meaning Representation AMRs are directed acyclic graph representations of sentence meaning (Banarescu et al., 2013), designed to capture the important meaning elements of a sentence while abstracting away from syntactic details This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/. 3694 Bill left for Paris (l / …

Content truncated to 500 words