Viable Dependency Parsing as Sequence Labeling

Proceedings of NAACL-HLT 2019, pages 717–723 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 717 Viable Dependency Parsing as Sequence Labeling Michalina Strzyz David Vilares Carlos G´omez-Rodr´ıguez Universidade da Coru˜na, CITIC FASTPARSE Lab, LyS Research Group, Departamento de Computaci´on Campus de Elvi˜na, s/n, 15071 A Coru˜na, Spain {michalina.strzyz,david.vilares,carlos.gomez}@udc.es Abstract We recast dependency parsing as a sequence labeling problem, exploring several encodings of dependency trees as labels. While depen- dency parsing by means of sequence labeling had been attempted in existing work, results suggested that the technique was impractical. We show instead that with a conventional BIL- STM-based model it is possible to obtain fast and accurate parsers. These parsers are con- ceptually simple, not needing traditional pars- ing algorithms or auxiliary structures. How- ever, experiments on the PTB and a sample of UD treebanks show that they provide a good speed-accuracy tradeoff, with results competi- tive with more complex approaches. 1 Introduction The application of neural architectures to syntac- tic parsing, and especially the ability of long short- term memories (LSTMs) to obtain context-aware feature representations (Hochreiter and Schmid- huber, 1997), has made it possible to parse natural language with conceptually simpler models than before. For example, in dependency parsing, the rich feature models with dozens of features used in transition-based approaches (Zhang and Nivre, 2011) can be simplified when using feedforward neural networks (Chen and Manning, 2014), and even more with BiLSTM architectures (Kiper- wasser and Goldberg, 2016), where in fact two positional features can suffice (Shi et al., 2017). Similarly, in graph-based approaches, Dozat and Manning (2017) have shown that an arc-factored model can achieve state-of-the-art accuracy, with- out the need for the higher-order features used in systems like (Koo and Collins, 2010). In the same way, neural feature representations have made it possible to relax the need for struc- tured representations. This is the case of sequence- to-sequence models that translate sentences into linearized trees, which were first applied to con- stituent (Vinyals et al., 2015) and later to depen- dency parsing (Wiseman and Rush, 2016; Zhang et al., 2017b; Li et al., 2018). Recently, G´omez- Rodr´ıguez and Vilares (2018) have shown that se- quence labeling models, where each word is asso- ciated with a label (thus simpler than sequence to sequence, where the mapping from input to output is not one to one) can learn constituent parsing. Contribution We show that sequence labeling is useful for dependency parsing, in contrast to previ- ous work (Spoustov´a and Spousta, 2010; Li et al., 2018). We explore four different encodings to rep- resent dependency trees for a sentence of length n as a set of n labels associated with its words. We then use these representations to perform depen- dency parsing with an off-the-shelf sequence la- beling model. The results show that we produce models with an excellent speed-accuracy tradeoff, without requiring any explicit parsing algorithm or auxiliary structure (e.g. stack or buffer). The source code is available at https://github. com/mstrise/dep2label 2 Parsing as …

Content truncated to 500 words