Generation from Abstract Meaning Representation using Tree Transducers

Proceedings of NAACL-HLT 2016, pages 731–739, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Generation from Abstract Meaning Representation using Tree Transducers Jeffrey Flanigan♠ Chris Dyer♠ Noah A. Smith♥ Jaime Carbonell♠ ♠School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA ♥Computer Science & Engineering, University of Washington, Seattle, WA, USA {jflanigan,cdyer,jgc}@cs.cmu.edu, nasmith@cs.washington.edu Abstract Language generation from purely semantic representations is a challenging task. This pa- per addresses generating English from the Ab- stract Meaning Representation (AMR), con- sisting of re-entrant graphs whose nodes are concepts and edges are relations. The new method is trained statistically from AMR- annotated English and consists of two major steps: (i) generating an appropriate spanning tree for the AMR, and (ii) applying tree-to- string transducers to generate English. The method relies on discriminative learning and an argument realization model to overcome data sparsity. Initial tests on held-out data show good promise despite the complexity of the task. The system is available open-source as part of JAMR at: http://github.com/jflanigan/jamr 1 Introduction We consider natural language generation from the Abstract Meaning Representation (AMR; Banarescu et al., 2013). AMR encodes the meaning of a sen- tence as a rooted, directed, acyclic graph, where con- cepts are nodes, and edges are relationships among the concepts. Because AMR models propositional meaning1 while abstracting away from surface syntactic re- alizations, and is designed with human annotation in mind, it suggests a separation of (i) engineering the application-specific propositions that need to be 1In essence, AMR handles semantic roles, entity types, within-sentence coreference, discourse connectives, modality, negation, and some other phenomena. communicated about from (ii) general-purpose re- alization details, modeled by a generator shareable across many applications. The latter is our focus here. Because any AMR graph has numerous valid re- alizations, and leaves underspecified many impor- tant details—including tense, number, definiteness, whether a concept should be referred to nominally or verbally, and more—transforming an AMR graph into an English sentence is a nontrivial problem. To our knowledge, our system is the first for gen- erating English from AMR. The approach is a sta- tistical natural language generation (NLG) system, trained discriminatively using sentences in the AMR bank (Banarescu et al., 2013). It first transforms the graph into a tree, then decodes into a string us- ing a weighted tree-to-string transducer and a lan- guage model (Graehl and Knight, 2004). The de- coder bears a strong similarity to state-of-the-art ma- chine translation systems (Koehn et al., 2007; Dyer et al., 2010), but with a rule extraction approach tai- lored to the NLG problem. 2 Overview Generation of English from AMR graphs is accom- plished as follows: the input graph is converted to a tree, which is input into the weighted intersection of a tree-to-string transducer (§4) with a language model. The output English sentence is the (ap- proximately) highest-scoring sentence according to a feature-rich discriminatively trained linear model. After discussing notation (§3), we describe our ap- proach in §4. The transducer’s rules are extracted from the limited AMR …

Content truncated to 500 words