Extracting Headless MWEs from Dependency Parse Trees: Parsing, Tagging, and Joint Modeling Approaches

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8780–8794 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8780 Extracting Headless MWEs from Dependency Parse Trees: Parsing, Tagging, and Joint Modeling Approaches Tianze Shi Cornell University tianze@cs.cornell.edu Lillian Lee Cornell University llee@cs.cornell.edu Abstract An interesting and frequent type of multi- word expression (MWE) is the headless MWE, for which there are no true internal syntactic dominance relations; examples include many named entities (“Wells Fargo”) and dates (“July 5, 2020”) as well as certain produc- tive constructions (“blow for blow”, “day after day”). Despite their special status and preva- lence, current dependency-annotation schemes require treating such flat structures as if they had internal syntactic heads, and most cur- rent parsers handle them in the same fash- ion as headed constructions. Meanwhile, out- side the context of parsing, taggers are typ- ically used for identifying MWEs, but tag- gers might benefit from structural information. We empirically compare these two common strategies—parsing and tagging—for predict- ing flat MWEs. Additionally, we propose an efficient joint decoding algorithm that com- bines scores from both strategies. Experimen- tal results on the MWE-Aware English Depen- dency Corpus and on six non-English depen- dency treebanks with frequent flat structures show that: (1) tagging is more accurate than parsing for identifying flat-structure MWEs, (2) our joint decoder reconciles the two differ- ent views and, for non-BERT features, leads to higher accuracies, and (3) most of the gains re- sult from feature sharing between the parsers and taggers. 1 Introduction Headless multi-word expressions (MWEs), includ- ing many named entities and certain productive constructions, are frequent in natural language and are important to NLP applications. In the con- text of dependency-based syntactic parsing, how- ever, they pose an interesting representational chal- lenge. Dependency-graph formalisms for syntactic structure represent lexical items as nodes and head- dominates-modifier/argument relations between Officials at Mellon Capital were unavailable for comment O O B I O O O O nsubj case nmod mwe_NNP xcomp case nmod Figure 1: Dependency tree from the MWE-Aware En- glish Dependency Corpus, imposing a “head” relation- ship between the words in the actually headless MWE Mellon Capital. Also shown are MWE BIO labels. lexical items as directed arcs on the correspond- ing pair of nodes. Most words can be assigned clear linguistically-motivated syntactic heads, but several frequently occurring phenomena do not eas- ily fit into this framework, including punctuation, coordinating conjunctions, and “flat”, or headless MWEs. While the proper treatment of headless constructions in dependency formalisms remains debated (Kahane et al., 2017; Gerdes et al., 2018), many well-known dependency treebanks handle MWEs by giving their component words a “default head”, which is not indicative of a true dominance relation, but rather as “a tree encoding of a flat structure without a syntactic head” (de Marneffe and Nivre, 2019, pg. 213). Fig. 1 shows an exam- ple: the headless MWE Mellon Capital has its first word, Mellon, marked as the “head” of Capital. Despite the special status …

Content truncated to 500 words