On Parsing as Tagging

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8884 - 8900 December 7-11, 2022 ©2022 Association for Computational Linguistics On Parsing as Tagging Afra Amini Ryan Cotterell {afra.amini, ryan.cotterell}@inf.ethz.ch Abstract There have been many proposals to reduce constituency parsing to tagging in the lit- erature. To better understand what these approaches have in common, we cast several existing proposals into a unifying pipeline consisting of three steps: linearization, learn- ing, and decoding. In particular, we show how to reduce tetratagging, a state-of-the-art constituency tagger, to shift–reduce parsing by performing a right-corner transformation on the grammar and making a specific independence assumption. Furthermore, we empirically evaluate our taxonomy of tagging pipelines with different choices of linearizers, learners, and decoders. Based on the results in English and a set of 8 typologically diverse languages, we conclude that the linearization of the derivation tree and its alignment with the input sequence is the most critical factor in achieving accurate taggers. https://github.com/rycolab/ parsing-as-tagging 1 Introduction The automatic syntactic analysis of natural language text is an important problem in natural language processing (NLP). Due to the ambiguity present in natural language, many parsers for natural language are statistical in nature, i.e., they provide a probability distribution over syntactic analyses for a given input sequence. A common design pattern for natural language parsers is to re-purpose tools from formal language theory that were often designed for compiler construction (Aho and Ullman, 1972), to construct probability distributions over derivation trees (syntactic analyses). Indeed, one of the most commonly deployed parsers in NLP is a statistical version of the classic shift–reduce parser, which has been widely applied in constituency parsing (Sagae and Lavie, 2005; Zhang and Clark, 2009; Zhu et al., 2013), dependency parsing (Fraser, 1989; Yamada She Linearizer + Aligner PRP V V N VP VP S She: (S), enjoys: (VP), reading: (VP), papers: enjoys reading papers She enjoys reading papers 1 2 Learner Decoder She: papers: enjoys: reading: 3 She PRP V V N VP VP S enjoys reading papers Figure 1: We subdivide parsing-as-tagging schemes into three steps: 1) linearization, 2) learning, and 3) decoding. The figure depicts how subdivision affects the parsing of the sentence She enjoys reading papers with pre-order linearization. and Matsumoto, 2003; Nivre, 2004), and even for the parsing of more exotic formalisms, e.g., CCG parsing (Zhang and Clark, 2011; Xu et al., 2014). Another relatively recent trend in statistical parsing is the idea of reducing parsing to tagging (Gómez-Rodríguez and Vilares, 2018; Strzyz et al., 2019; Vilares et al., 2019; Vacareanu et al., 2020; Gómez-Rodríguez et al., 2020; Kitaev and Klein, 2020). There are a number of motivations behind this reduction. For instance, Kitaev and Klein (2020) motivate tetratagging, their novel constituency parsing-as-tagging scheme, by its ability to parallelize the computation during training and to minimize task-specific modeling. The proposed taggers in the literature are argued to 8884 exhibit the right balance in the trade-off between accuracy and efficiency. Nevertheless, there are several ways to …

Content truncated to 500 words