Weakly Supervised Word Segmentation for Computational Language Documentation

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7385 - 7398 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Weakly Supervised Word Segmentation for Computational Language Documentation Shu Okabe Univ. Paris-Saclay & CNRS LISN, rue John von Neumann 91403 Orsay, France shu.okabe@lisn.fr Laurent Besacier Naver Labs Europe 6 chem. de Maupertuis 38240 Meylan, France laurent.besacier@naverlabs.com François Yvon Univ. Paris-Saclay & CNRS LISN, rue John von Neumann 91403 Orsay, France francois.yvon@limsi.fr Abstract Word and morpheme segmentation are funda- mental steps of language documentation as they allow to discover lexical units in a lan- guage for which the lexicon is unknown. How- ever, in most language documentation scenar- ios, linguists do not start from a blank page: they may already have a pre-existing dictio- nary or have initiated manual segmentation of a small part of their data. This paper studies how such a weak supervision can be taken ad- vantage of in Bayesian non-parametric models of segmentation. Our experiments on two very low resource languages (Mboshi and Japhug), whose documentation is still in progress, show that weak supervision can be beneficial to the segmentation quality. In addition, we investi- gate an incremental learning scenario where manual segmentations are provided in a se- quential manner. This work opens the way for interactive annotation tools for documen- tary linguists. 1 Introduction Recent years have witnessed a blooming of re- search aimed at applying language technologies (LTs) to “under-resourced languages”.1 Such stud- ies have been mostly motivated on three main grounds (not necessarily mutually exclusive): (a) to develop tools that could speed up the work of field linguists collecting and annotating recordings for these languages; (b) to provide linguistic communi- ties with LTs that are necessary in an increasingly digitalised world, e.g. to interact with smartphones or computers in their own language and communi- cate with speakers of other languages; (c) to chal- lenge existing machine-learning techniques in very low resource settings, where hardly any resource (dictionary, corpus, grammar) is available. 1Acknowledged by workshop series such as “Spoken Lan- guages Technologies for Under-resourced languages (SLTU), “Collaboration and Computing for Under-Resourced Lan- guages” (CCURL) and “Computational Methods in the Study of Endangered Languages” (ComputEL) inter alia. Those objectives are thoroughly discussed in a recent position paper (Bird, 2020) who notices, among other things, that objective (c) (training lan- guage processing tools with zero resource) is ques- tionable in the context of language documentation works which can often rely on some pre-existing knowledge, such as a word list, or information from related languages. Accordingly, this paper explores ways to make the best of prior resources and im- prove the effectiveness of unsupervised language analysis techniques for the purpose of linguistic documentation. Our main objective is to develop tools that will effectively assist field linguists in their documentary tasks (objective (a)). We focus on segmentation tasks, which aim to automatically identify meaningful units in an unsegmented pho- netic or orthographic string (Johnson, 2008; Doyle and Levy, 2013; Eskander et …

Content truncated to 500 words