UC Davis at SemEval-2019 Task 1: DAG Semantic Parsing with Attention-based Decoder

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 119–124 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 119 UC Davis at SemEval-2019 Task 1: DAG Semantic Parsing with Attention-based Decoder Dian Yu University of California, Davis dianyu@ucdavis.edu Kenji Sagae University of California, Davis sagae@ucdavis.edu Abstract We present a simple and accurate model for semantic parsing with UCCA as our submis- sion for SemEval 2019 Task 1. We propose an encoder-decoder model that maps strings to di- rected acyclic graphs. Unlike many transition- based approaches, our approach does not use a state representation, and unlike graph-based parsers, it does not score graphs directly. Instead, we encode input sentences with a bidirectional-LSTM, and decode with self- attention to build a graph structure. Results show that our parser is simple and effective for semantic parsing with reentrancy and discon- tinuous structures. 1 Introduction Semantic parsing aims to capture structural re- lationships between input strings and graph rep- resentations of sentence meaning, going beyond concerns of surface word order, phrases and re- lationships. The focus on meaning rather than surface relations often requires the use of reen- trant nodes and discontinuous structures. Uni- versal Conceptual Cognitive Annotation (UCCA) (Abend and Rappoport, 2013) is designed to sup- port semantic parsing with mappings between sen- tences and their corresponding meanings in a framework intended to be applicable across lan- guages. SemEval 2019 Task 1 (Hershcovich et al., 2018b, 2019) focuses on semantic parsing of texts into graphs consisting of terminal nodes that repre- sent words, non-terminal nodes that represent in- ternal structure, and labeled edges representing re- lationships between nodes (e.g. participant, cen- ter, linker, adverbial, elaborator), according to the UCCA scheme. Annotated datasets are pro- vided, and participants are evaluated in four set- tings: English with domain-specific data, English with out-of-domain data, German with domain- specific data, and French with only development and test data, but no training data. Additionally, there are open and closed tracks, where the use of additional resources is and is not allowed, respec- tively. Our entry in the task is limited to the closed track and the first setting, domain-specific English using the Wiki corpus, where the relatively small dataset (4113 sentences for training, 514 for devel- opment, and 515 for testing) consists of annotated sentences from English Wikipedia. Our model follows the encoder-decoder archi- tecture commonly used in state-of-the-art neural parsing models (Kitaev and Klein, 2018; Kiper- wasser and Goldberg, 2016b; Cross and Huang, 2016; Chen and Manning, 2014). However, we propose a very simple decoder architecture that relies only on a recursive attention mechanism of the encoded latent representation. In other words, the decoder does not require state encod- ing and model-optimal inference whatsoever. Our novel model achieved a macro-averaged F1-score of 0.753 in labeled primary edges and 0.864 in unlabeled primary edge prediction on the test set. The results confirm the suitability of our proposed model to the semantic parsing task. 2 Related work Leveraging parallels between UCCA and known approaches …

Content truncated to 500 words