An Inflectional Database for Gitksan

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 6597–6606 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 6597 An Inflectional Database for Gitksan Bruce OliverΥ, Clarissa ForbesΦ, Changbing YangΥ, Farhan SamirΥ, Edith CoatesΥ, Garrett NicolaiΥ, Miikka SilfverbergΥ ΥUniversity of British Columbia, ΦIndependent bruce.oliver@ubc.ca, forbesc@alumni.ubc.ca, cyang33@student.ubc.ca, fsamir@mail.ubc.ca icoates1@mail.ubc.ca, garrett.nicolai@mail.ubc.ca, msilfver@mail.ubc.ca Abstract This paper presents a new inflectional resource for Gitksan, a low-resource Indigenous language of Canada. We use Gitksan data in interlinear glossed format, stemming from language documentation efforts, to build a database of partial inflection tables. We then enrich this morphological resource by filling in blank slots in the partial inflection tables using neural transformer reinflection models. We extend the training data for our transformer reinflection models using two data augmentation techniques: data hallucination and back-translation. Experimental results demonstrate substantial improvements from data augmentation, with data hallucination delivering particularly impressive gains. We also release reinflection models for Gitksan. Keywords: Gitksan, computational morphology, low-resource, reinflection 1. Introduction In recent years, research on computational morphol- ogy has enjoyed access to an ever-increasing variety of annotated data through the Universal Morphology (UniMorph) database (McCarthy et al., 2020)—a col- lection of inflection tables which now covers hundreds of languages (see Figure 1 for an example). Most of the data in UniMorph stems from Wiktionary and is auto- matically converted into UniMorph format. Wiktionar- ies are, however, available only for a limited number of languages. In this paper, we explore an alternative route for generation of high-quality paradigm-level in- flectional data. Our starting-point is interlinear glossed text (IGT) for the Tsimshianic language Gitksan, gen- erated in language documentation efforts. We convert this text into a paradigm-level inflectional resource us- ing a combination of rule-based methods and neural morphological reinflection (Cotterell et al., 2016). IGT is a semi-structured representation format which is commonly employed in language documentation. It consists of transcribed utterances paired with morpho- logical segmentation information, morphosyntactic an- notations, and translations into a meta language (En- glish in our case). See Figure 2 for a Gitksan example. IGT is a useful starting point for construction of mor- phological databases because of its extensive morpho- logical annotation. Conversion of IGT into inflection tables entails many design decisions. For example, due to the under- documented status of Gitksan, it is not clear which mor- phological processes constitute inflection and which are more properly classified as derivation. Section 5 presents our criteria for distinguishing between Gitk- san inflection and derivation along with the conversion pipeline from IGT to inflection tables. Our paper is accompanied by the first database of in- flection tables for Gitksan. Since not all inflected forms of every lexeme are attested in our underlying IGT re- source, we additionally release neural reinflection mod- els for filling in missing forms in the tables.1. The final inflectional database has a multitude of applications: it can be used for construction of language learning resources, experiments in computational morphology and training of automatic systems for glossing, which can …

Content truncated to 500 words