CUNY-PKU Parser at SemEval-2019 Task 1: Cross-Lingual Semantic Parsing with UCCA

CUNY-PKU Parser at SemEval-2019 Task 1: Cross-lingual Semantic Parsing with UCCA Weimin Lyu†, Sheng Huang‡, Abdul Rafae Khan†, Shengqiang Zhang‡ , Weiwei Sun‡, Jia Xu†⋄ Computer Science Department, †Graduate Center and ⋄Hunter College, City University of New York {wlyu,akhan4}@gradcenter.cuny.edu, Jia.Xu@hunter.cuny.edu ‡Institute of Computer Science and Technology, Peking University {huangsheng,ws,sq.zhang}@pku.edu.cn Abstract This paper describes the systems of the CUNY-PKU team in “SemEval 2019 Task 1: Cross-lingual Semantic Parsing with UCCA”1. We introduce a novel method by applying a cascaded MLP and BiLSTM model. Then, we ensemble multiple system-outputs by repars- ing. Our system won the second places in German-20K-Closed track, and third place in English-20K-Closed track. 1 Introduction We participate in Cross-lingual Semantic Parsing at SemEval 2019, and our submission systems are based on TUPA (Hershcovich et al., 2017a, 2018). A shared task summary paper (Hershcovich et al., 2019) by competition organizers summaries the results. We built three single parser using BiLSTM (Bi- directional LSTM) and Multi-Layer Perceptron (MLP) with TUPA (Hershcovich et al., 2017a, 2018). Most importantly, we introduce a new training method Cascaded BiLSTM by first pre- training the BiLSTM model and then training an- other MLP model based on pre-trained BiLSTM model. The cascaded BiLSTM parser enhances the parsing accuracy on all tasks. We also com- plete a Self-Attentive Constituency Parser (Kitaev and Klein, 2018a,b) as comparison. Finally, we ensemble different parsers with a reparsing strat- egy (Sagae and Lavie, 2006). In particular, we in- troduce an algorithm based on dynamic program- ming to perform inference for the UCCA repre- sentation. This decoder can also be utilized as a core engine for a single parser. We will describe our systems in detail, includ- ing three single parsers in Section 2 and a voter in 1https://competitions.codalab.org/ competitions/19160 Section 3. We focus on two novel technical con- tributions: the Cascaded BiLSTM model and the Reparsing strategy. In Section 4 we will present experimental setup and results. 2 Single Parsers 2.1 TUPA Parsers The TUPA parser (Hershcovich et al., 2017a) builds on discontinuous constituency and depen- dency graph parsing and makes some improve- ments especially for the UCCA representation. The English parsing is based on Hershcovich et al. (2017a), while French and German parsing is based on Hershcovich et al. (2018). It has been shown that the choice of model plays an important role in transition-based pars- ing (Hershcovich et al., 2017b). For TUPA, we built parsers with different models: MLP, BiL- STM, and also train MLP based on BiLSTM, viz. Cascaded BiLSTM. The three single parsers are described as the following: The MLP parser (Hershcovich et al., 2017b) applies a feedforward neural network with dense embedding features to predict optimal transitions given particular parser states. This parser adopts a similar architecture to Chen and Manning (2014). The BiLSTM parser (Hershcovich et al., 2018) applies a bidirectional LSTM to learn con- textualized vector-based representations for words that are then utilized for encoding a parser state, similarly to Kiperwasser and Goldberg (2016). The red box in Figure 1 shows the architecture of …

Content truncated to 500 words