Spoken Language Treebanks in Universal Dependencies: an Overview

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1798–1806 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1798 Spoken Language Treebanks in Universal Dependencies: an Overview Kaja Dobrovoljc Faculty of Arts, University of Ljubljana Jožef Stefan Institute, Ljubljana, Slovenia kaja.dobrovoljc@ff.uni-lj.si Abstract Given the benefits of syntactically annotated collections of transcribed speech in spoken language research and applications, many spoken language treebanks have been developed in the last decades, with divergent annotation schemes posing important limitations to cross-resource explorations, such as comparing data across languages, grammatical frameworks, and language domains. As a consequence, there has been a growing number of spoken language treebanks adopting the Universal Dependencies (UD) annotation scheme, aimed at cross-linguistically consistent morphosyntactic annotation. In view of the non-central role of spoken language data within the scheme and with little in-domain consolidation to date, this paper presents a comparative overview of spoken language treebanks in UD to support cross-treebank data explorations on the one hand, and encourage further treebank harmonization on the other. Our results show that the spoken language treebanks differ considerably with respect to the inventory and the format of transcribed phenomena, as well as the principles adopted in their morphosyntactic annotation. This is particularly true for the dependency annotation of speech disfluencies, where conflicting data annotations suggest an underspecification of the guidelines pertaining to speech repairs in general and the reparandum dependency relation in particular. Keywords: Universal Dependencies, treebanks, speech transcriptions, morphosyntactic annotation, dependency syntax, disfluencies, speech repairs 1. Introduction Spoken language treebanks, i.e. syntactically an- notated collections of transcribed speech, represent one of the fundamental language resources for spo- ken language processing tasks, such as syntactic parsing (Caines et al., 2017; Dobrovoljc and Mart- inc, 2018; Braggaar and van der Goot, 2021; Liu and Prud’hommeaux, 2021) and information retrieval (Davidson et al., 2019; Liu et al., 2021). Spoken lan- guage treebanks are equally important in linguistic re- search in general, not only due to speech being the primary and prevailing mode of human communica- tion that exhibits several idiosyncrasies in comparison to writing (Biber et al., 2010; Carter and McCarthy, 2015), but also due to the fact that many of the world languages have no written form at all. Many spoken language treebanks have been created since the pivotal work on the Switchboard section of the Penn Treebank (Marcus et al., 1993), such as the multilingual Tübingen (Hinrichs et al., 2000) and CHILDES (Sagae et al., 2004; MacWhinney, 2014) treebank collections, not to mention the number of treebanks developed for individual languages. For most, customized treebank-specific syntactic annota- tion schemes were developed, adopting divergent ap- proaches to annotating syntactic phenomena in gen- eral and speech-specific phenomena in particular, pos- ing important limitations to various kinds of cross- resource explorations, such as data comparisons across languages, grammatical frameworks or language vari- eties. In line with the growing need for spoken data standard- ization and consolidation, there has also been an in- creasing number of spoken language …

Content truncated to 500 words