RIGOTRIO at SemEval-2017 Task 9: Combining Machine Learning and Grammar Engineering for AMR Parsing and Generation

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 924–928, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics RIGOTRIO at SemEval-2017 Task 9: Combining Machine Learning and Grammar Engineering for AMR Parsing and Generation Normunds Gruzitis1, Didzis Gosko2 and Guntis Barzdins1 1University of Latvia, IMCS / Rainis blvd. 29, Riga, Latvia normunds.gruzitis@lumii.lv, guntis.barzdins@lumii.lv 2LETA / Marija street 2, Riga, Latvia didzis.gosko@leta.lv Abstract By addressing both text-to-AMR parsing and AMR-to-text generation, SemEval- 2017 Task 9 established AMR as a pow- erful semantic interlingua. We strengthen the interlingual aspect of AMR by apply- ing the multilingual Grammatical Frame- work (GF) for AMR-to-text generation. Our current rule-based GF approach com- pletely covered only 12.3% of the test AMRs, therefore we combined it with state-of-the-art JAMR Generator to see if the combination increases or decreases the overall performance. The combined sys- tem achieved the automatic BLEU score of 18.82 and the human Trueskill score of 107.2, to be compared to the plain JAMR Generator results. As for AMR parsing, we added NER extensions to our SemEval-2016 general-domain AMR parser to handle the biomedical genre, rich in organic compound names, achieving Smatch F1=54.0%. 1 Introduction AMR (Banarescu et al., 2013) as a sentence-level semantic representation is evolving towards inter- lingua at SemEval-2017 Task 9 on Abstract Mean- ing Representation Parsing and Generation (May and Priyadarshi, 2017). The challenge was to im- prove over state-of-the-art systems for both text- to-AMR parsing (Barzdins and Gosko, 2016) and AMR-to-text generation (Flanigan et al., 2016). AMR parsing subtask this year focused on spe- cific genre of Biomedical scientific articles regard- ing cancer pathway discovery. Such texts are chal- lenging to existing AMR parsers because they are rich in organic compound names with types “en- zyme”, “aminoacid”, etc. not recognized by com- mon NER tools that are often restricted to types “person”, “organization”, “location”, etc. The paper starts with NER extensions used for the Biomedical AMR parsing subtask, followed by a novel approach of using Grammatical Frame- work for AMR generation, and concludes with a brief analysis of our SemEval results. 2 Text-to-AMR parsing Only two adaptations to the AMR parser from SemEval-2016 (Barzdins and Gosko, 2016) were implemented: it was retrained on the union of LDC2015E86, LDC2016E25, LDC2016E33 and Bio AMR Corpus, and a gazetteer was added to extend the NER coverage to organic compound names found in the Bio AMR Corpus (e.g. “B- Raf enzyme”, “dabrafenib small-molecule”, etc.). The gazetteer was generalized w.r.t. numbers used in the names. Although we achieved an above average Smatch score (54.0% versus 53.6%) in the preliminary official scoring, the ablation metrics show that we scored below average for named entities (46.0% versus 55.8%) and wikification (0% versus 33.0%). Since we used gazetteers extracted from the training data for both named entities and wik- ification, this suggests that external data sources should have been used instead. 3 AMR-to-text generation Our approach to text generation from AMR graphs stems from a recent feasibility study (Gruzitis and Barzdins, 2016) …

Content truncated to 500 words