RIGA at SemEval-2016 Task 8: Impact of Smatch Extensions and Character-Level Neural Translation on AMR Parsing Accuracy

Two extensions to the AMR smatch scoring script are presented. The first\nextension com-bines the smatch scoring script with the C6.0 rule-based\nclassifier to produce a human-readable report on the error patterns frequency\nobserved in the scored AMR graphs. This first extension results in 4% gain over\nthe state-of-art CAMR baseline parser by adding to it a manually crafted\nwrapper fixing the identified CAMR parser errors. The second extension combines\na per-sentence smatch with an en-semble method for selecting the best AMR graph\namong the set of AMR graphs for the same sentence. This second modification\nau-tomatically yields further 0.4% gain when ap-plied to outputs of two\nnondeterministic AMR parsers: a CAMR+wrapper parser and a novel character-level\nneural translation AMR parser. For AMR parsing task the character-level neural\ntranslation attains surprising 7% gain over the carefully optimized word-level\nneural translation. Overall, we achieve smatch F1=62% on the SemEval-2016\nofficial scor-ing set and F1=67% on the LDC2015E86 test set.\n