Abstract Meaning Representation Parsing using LSTM Recurrent Neural Networks

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 463–472 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1043 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 463–472 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1043 Abstract Meaning Representation Parsing using LSTM Recurrent Neural Networks William R. Foland Jr. Department of Computer Science University of Colorado Boulder, CO 80309 William.Foland@colorado.edu James H. Martin Department of Computer Science and Institute of Cognitive Science University of Colorado Boulder, CO 80309 James.Martin@colorado.edu Abstract We present a system which parses sen- tences into Abstract Meaning Representa- tions, improving state-of-the-art results for this task by more than 5%. AMR graphs represent semantic content using linguistic properties such as semantic roles, coref- erence, negation, and more. The AMR parser does not rely on a syntactic pre- parse, or heavily engineered features, and uses five recurrent neural networks as the key architectural components for inferring AMR graphs. 1 Introduction Semantic analysis is the process of extracting meaning from text, revealing key ideas such as ”who did what to whom, when, how, and where?”, and is considered to be one of the most complex tasks in natural language processing. Historically, an important consideration has been the definition of the output of the task - how can the concepts in a sentence be captured in a general, consistent and expressive manner that facilitates downstream semantic processing? Over the years many for- malisms have been proposed as suitable target rep- resentations including variants of first order logic, semantic networks, and frame-based slot-filler no- tations. Such representations have found a place in many semantic applications but there is no clear consensus as to the best representation. However, with the rise of supervised machine learning tech- niques, a new requirement has come to the fore: the ability of human annotators to quickly and reli- ably generate semantic representations as training data. Abstract Meaning Representation (AMR) (Ba- narescu et al., 2012)1 was developed to provide 1http://amr.isi.edu/language.html a computationally useful and expressive repre- sentation that could be reliably generated by hu- man annotators. Sentence meanings in AMR are represented in the form of graphs consisting of concepts (nodes) connected by labeled relations (edges). AMR graphs include a number of tra- ditional NLP representations including named en- tities (Nadeau and Sekine, 2007), word senses (Banerjee and Pedersen, 2002), coreference rela- tions, and predicate-argument structures (Kings- bury and Palmer, 2002; Palmer et al., 2005). More recent innovations include wikification of named entities and normalization of temporal expressions (Verhagen et al., 2010; Str¨otgen and Gertz, 2010). (2016) provides an insightful discussion of the re- lationship between AMR and other formal repre- sentations including first order logic. The process of creating AMR’s for sentences is called AMR Parsing and was first introduced in (Flanigan et al., 2014). A key factor driving the de- velopment of AMR systems has been the increas- ing availability of training resources in the …

Content truncated to 500 words