Unmasking the Myth of Effortless Big Data - Making an Open Source Multi-lingual Infrastructure and Building Language Resources from Scratch

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1167–1177 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1167 Unmasking the Myth of Effortless Big Data — Making an Open Source Multilingual Infrastructure and Building Language Resources from Scratch Linda Wiechetek, Katri Hiovain-Asikainen, Inga Lill Sigga Mikkelsen, Sjur N. Moshagen, Flammie A. Pirinen, Trond Trosterud, Børre Gaup Department of Language and Culture, UiT the Arctic University of Norway giellalt@uit.no Abstract Machine learning (ML) approaches have dominated Natural Language Processing (NLP) during the last two decades. From machine translation and speech technology, machine learning tools are now also in use for spellchecking and grammar checking, with a blurry distinction between the two. We unmask the myth of effortless big data by illuminating the efforts and time that lay behind building a multi-purpose corpus with regard to collecting, marking up and building from scratch. We also discuss what kind of language technology tools minority language communities actually need, and to what extent the dominating paradigm has been able to deliver these tools. In this context we present our alternative to corpus-based language technology – knowledge-based language technology – and we show how this approach can provide language technology solutions for languages being outside the reach of machine learning procedures. We present a stable and mature infrastructure (GiellaLT) containing more than hundred languages and building a number of language technology tools that are useful for language communities. Keywords: infrastructure, corpus, text processing, minority languages, finite state technology, knowledge-based nlp, grammar checking, TTS, ASR, speech technology, spellchecking 1. Introduction During the last two decades, i.e. early 2000s, machine learning approaches have dominated the field of natural language processing (NLP). The philosophy has been to have machines learn behaviour from large corpora, thereby offering speech technology, machine transla- tion and spelling and grammar checking. For at least 3/4 of the world’s languages, this is bad news. They have less than 20,000 speakers, and even with a school system teaching mother tongue literacy and a language policy encouraging publishing, such small populations would have a hard time producing enough text to be able to model the language in a re- liable way using the dominating NLP paradigm. On top of that, weak literary traditions give rise to corpora too unreliable to be able to function as a model for NLP tools. From the reactions we get on our work in the cur- rent academic NLP community contexts, we experi- ence an absence of understanding for why we choose to work rule-based, despite our explanations of unavail- able noise-free data. There seems to be the idea that big data comes for free. In addition, since most of the work within the dominating methodology is done on morpho- logically simple languages, there seems to be a general lack of understanding of the methodologies and tech- nologies needed to work efficiently with polysynthetic and other morphologically complex languages. This article has a clear answer to the big data question — …

Content truncated to 500 words