Auxiliary tasks to boost Biaffine Semantic Dependency Parsing

Findings of the Association for Computational Linguistics: ACL 2022, pages 2422 - 2429 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Auxiliary Tasks to Boost Biaffine Semantic Dependency Parsing Marie Candito LLF, Université Paris Cité - CNRS, Paris, France marie.candito@u-paris.fr Abstract The biaffine parser of Dozat and Manning (2017) was successfully extended to semantic dependency parsing (SDP) (Dozat and Man- ning, 2018). Its performance on graphs is sur- prisingly high given that, without the constraint of producing a tree, all arcs for a given sen- tence are predicted independently from each other (modulo a shared representation of to- kens). To circumvent such an independence of decision, while retaining the O(n2) complex- ity and highly parallelizable architecture, we propose to use simple auxiliary tasks that intro- duce some form of interdependence between arcs. Experiments on the three English acyclic datasets of SemEval 2015 task 18 (Oepen et al., 2015), and on French deep syntactic cyclic graphs (Ribeyre et al., 2014) show modest but systematic performance gains on a near state-of- the-art baseline using transformer-based con- textualized representations. This provides a simple and robust method to boost SDP perfor- mance. 1 Introduction and related work Semantic dependency parsing is the task of pro- ducing a dependency graph for a sentence. De- pending on the datasets, these dependencies may correspond to predicate-argument relations, with labels numbering semantic arguments (as in Fig- ure 1-top) or dependencies with intermediate status between syntax and semantics, with labels being canonical grammatical functions that normalize syntactic alternations (e.g. in Figure 1-bottom, the clitic l’ (him) is the canonical object of the passive verb form sollicité (solicited)). If one views each dependency as a decision to make, both dependency parsing (DP, outputing a syntactic tree) and semantic dependency parsing (SDP) are known to exhibit high interdependence of decisions. For instance in DP, when parsing a question answering machine, choosing machine as root is linguistically coherent with the machine She went back and spoke to the desk clerk. ARG1 ARG1 ROOT LOC AND_C ARG1 BV COMPOUND ARG2 Cela l’ a habitué à être très sollicité. This him has accustomed to be very solicited. suj obj mod obj Figure 1: Top: English Semantic graph in the DM for- mat, as part of the SemEval2015-Task18 dataset (Oepen et al., 2015). Bottom: French Deep syntactic graph as defined by Candito et al. (2014). → answering → question analysis only, whereas (wrongly) choosing question as root is syntacti- cally coherent with the question → answering → machine analysis. In DP though, the interdependence between arcs is partially solved by the tree constraint: choosing one head for a given token amounts to ruling out all other heads. This structural interdependence is absent in SDP. Complex structural, lexical and semantic factors control whether a given dependent should be attached to zero, one or several heads. Several approaches exist in the literature to cap- ture interdependence of arcs in SDP, which often derive from proposals made for DP. One is to use a higher-order graph-based parser. Wang …

Content truncated to 500 words