Don’t Parse, Choose Spans! Continuous and Discontinuous Constituency Parsing via Autoregressive Span Selection

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8420–8433 July 9-14, 2023 ©2023 Association for Computational Linguistics Don’t Parse, Choose Spans! Continuous and Discontinuous Constituency Parsing via Autoregressive Span Selection Songlin Yang, Kewei Tu∗ School of Information Science and Technology, ShanghaiTech University Shanghai Engineering Research Center of Intelligent Vision and Imaging yangsl66@mit.edu, tukw@shanghaitech.edu.cn Abstract We present a simple and unified approach for both continuous and discontinuous con- stituency parsing via autoregressive span selec- tion. Constituency parsing aims to produce a set of non-crossing spans so that they can form a constituency parse tree. We sort gold spans in a predefined order and train a pointer network to autoregressively select spans by that order. To deal with a discontinuous span, we consec- utively select its subspans from left to right, label all but the last subspans with a special discontinuous label, and label the last subspan with the whole discontinuous span’s label. We use a simple heuristic to output valid trees from selected spans so that our approach is able to predict all possible continuous and discontinu- ous constituency trees without sacrificing data coverage and without the need to use expensive chart-based parsing algorithms. Extensive ex- periments show that our model achieves state- of-the-art or competitive performance on all benchmarks of continuous and discontinuous constituency parsing .1 1 Introduction Constituency parsing is a fundamental task in nat- ural language processing, having many applica- tions in downstream tasks such as language model- ing (Sartran et al., 2022) and machine translation (Akoury et al., 2019). Continuous constituency parsing is not capable of capturing discontinuous language phenomena, such as wh-movements and extraposition, which are deemed unavoidable in syntactic analysis (McCawley, 1982; Bunt et al., 1987; Müller, 2004). Discontinuous constituency parsing enables modeling these phenomena by al- lowing each constituent node governing nonadja- cent strings in the yields, resulting in discontinuous parse trees as exemplified in Fig. 1(b). In this work ∗Corresponding author 1Code is available at https://github.com/ sustcsonglin/disco-pointer. NP VP NP NP VP NP NP NP VP Span representation: (0, 2, NP), (0, 8, S), (2, 8, VP), (3, 5, NP), (3, 8, NP), (5, 8, PP), (7, 8, NP) (a) 0 The spoon NP PP NP S a with children ate the cake 1 2 3 4 5 6 7 8 VP today issue the on scheduled is hearing A NP NP PP NP NP VP Span representation: (0, 2, NP), (5, 7, NP), (4, 7, PP), (b) (7, 8, NP), (0, 8, VP), ((0, 2), (4, 7), NP), ((0,2), (3, 8), VP) 0 1 2 3 4 5 6 7 8 NP VP NP Figure 1: Examples of (a) continuous and (b) discontin- uous constituency parse trees with their span representa- tions. Continuous and discontinuous nodes are colored in blue and orange, respectively. In (b) all discontinuous nodes are of fan-out two. we study both continuous and discontinuous con- stituency parsing under a unified framework. Both continuous and discontinuous parsing can be framed …

Content truncated to 500 words