AMR Parsing with an Incremental Joint Model

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 680–689, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics AMR Parsing with an Incremental Joint Model Junsheng Zhou†, Feiyu Xu‡, Hans Uszkoreit‡, Weiguang Qu†, Ran Li† and Yanhui Gu† † Language Information Processing and Social Computing Lab School of Computer Science and Technology, Nanjing Normal University, China {zhoujs, wgqu, gu}@njnu.edu.cn, liran3277@sina.com ‡ Language Technology Lab, DFKI, Germany {feiyu, uszkoreit}@dfki.de Abstract To alleviate the error propagation in the tra- ditional pipelined models for Abstract Mean- ing Representation (AMR) parsing, we for- mulate AMR parsing as a joint task that per- forms the two subtasks: concept identification and relation identification simultaneously. To this end, we first develop a novel component- wise beam search algorithm for relation iden- tification in an incremental fashion, and then incorporate the decoder into a unified frame- work based on multiple-beam search, which allows for the bi-directional information flow between the two subtasks in a single incre- mental model. Experiments on the public datasets demonstrate that our joint model sig- nificantly outperforms the previous pipelined counterparts, and also achieves better or com- parable performance than other approaches to AMR parsing, without utilizing external se- mantic resources. 1 Introduction Producing semantic representations of text is moti- vated not only by theoretical considerations but also by the hypothesis that semantics can be used to im- prove many natural language tasks such as question answering, textual entailment and machine transla- tion. Banarescu et al. (2013) described a semantics bank of English sentences paired with their logical meanings, written in Abstract Meaning Representa- tion (AMR), which is rapidly emerging as an impor- tant practical form of structured sentence semantics. Recently, some literatures reported some promising applications of AMR. Pan et al. (2015) presented an unsupervised entity linking system with AMR, achieving the performance comparable to the super- vised state-of-the-art. Liu et al. (2015) demonstrated a novel abstractive summarization framework driven by the AMR graph that shows promising results. Garg et al. (2016) showed that AMR can signifi- cantly improve the accuracy of a biomolecular in- teraction extraction system compared to only using surface- and syntax-based features. Mitra and Baral (2016) presented a question-answering system by exploiting the AMR representation, obtaining good performance. Automatic AMR parsing is still in a nascent stage. Flanigan et al. (2014) built the first AMR parser, JAMR, based on a pipelined approach, which breaks down the whole task into two separate sub- tasks: concept identification and relation identifica- tion. Considering that node generation is an impor- tant limiting factor in AMR parsing, Werling et al. (2015) proposed an improved approach to the con- cept identification subtask by using a simple clas- sifier over actions which generate these subgraphs. However, the overall architecture is still based on the pipelined model. As a common drawback of the staged architec- ture, errors in upstream component are often com- pounded and propagated to the downstream predic- tion. The downstream components, however, can- not impact earlier decision. For …

Content truncated to 500 words