Sheffield at SemEval-2017 Task 9: Transition-based language generation from AMR.

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 586–591, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics Sheffield at SemEval-2017 Task 9: Transition-based language generation from AMR. Gerasimos Lampouras Department of Computer Science University of Sheffield, UK g.lampouras@sheffield.ac.uk Andreas Vlachos Department of Computer Science University of Sheffield, UK a.vlachos@sheffield.ac.uk Abstract This paper describes the submission by the University of Sheffield to the SemEval 2017 Abstract Meaning Representation Parsing and Generation task (SemEval 2017 Task 9, Subtask 2). We cast language generation from AMR as a sequence of actions (e.g., insert/remove/rename edges and nodes) that progressively transform the AMR graph into a dependency parse tree. This transition-based approach re- lies on the fact that an AMR graph can be considered structurally similar to a depen- dency tree, with a focus on content rather than function words. An added benefit to this approach is the greater amount of data we can take advantage of to train the parse- to-text linearizer. Our submitted run on the test data achieved a BLEU score of 3.32 and a Trueskill score of -2.204 on auto- matic and human evaluation respectively. 1 Introduction Abstract meaning representation (AMR) is a for- malism representing the meaning of a sentence (or multiple sentences) as a directed, acyclic graph, where each node represents a concept, and each edge represents a relation between concepts (Ba- narescu et al., 2013). Natural language generation (NLG) from AMRs introduces challenges, as AMR abstracts away from syntactic structure, function words, or inflections. Flanigan et al. (2016) were the first work to perform NLG from AMR; they used a weighted combination of a tree-to-string transducer and a language model to transform the AMR graph into English. Later work by Song et al. (2016) proposed segmenting the AMR graph into fragments and generating subphrases from them, using a set of subgraph-to-string rules. They then cast the problem of ordering these subphrases as a travelling salesman problem. Pourdamghani et al. (2016) suggested linearizing the AMR graph using a maximum entropy classifier. The linearization is then used as input to a phrase-based machine translation system, to produce the final sentence. Our submission to SemEval task 9 on AMR-to- English Generation is based on inverting previous work on transition-based parsers (Goodman et al., 2016a,b), which was in turn based on the previous work of Wang et al. (2015). Beyond inverting the transition from AMR graph to dependency tree, our system also separates the transition in three passes. Briefly, during the first pass we convert the AMR concepts into content words, during the second pass the structure of the tree is modified (e.g. by inserting, deleting, and moving nodes and edges), while in the third pass missing function words are inserted, and existing words realized in their final form. To form a natural language sentence, the de- pendency tree needs only to be linearized; we note that this is not part of the transition, but should be considered a separate post-processing step. We train a …

Content truncated to 500 words