Unsupervised Data Augmentation for Less-Resourced Languages with no Standardized Spelling

Proceedings of Recent Advances in Natural Language Processing, pages 776–784, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_090 776 Unsupervised Data Augmentation for Less-Resourced Languages with no Standardized Spelling Alice Millour and Kar¨en Fort Sorbonne Universit´e / STIH - EA 4509 28, rue Serpente, 75006 Paris, France alice.millour@etu.sorbonne-universite.fr, karen.fort@sorbonne-universite.fr Abstract Non-standardized languages are a chal- lenge to the construction of representative linguistic resources and to the develop- ment of efficient natural language process- ing tools: when spelling is not determined by a consensual norm, a multiplicity of al- ternative written forms can be encountered for a given word, inducing a large propor- tion of out-of-vocabulary words. To embrace this diversity, we propose a methodology based on crowdsourcing al- ternative spellings from which variation rules are automatically extracted. The rules are further used to match out- of-vocabulary words with one of their spelling variants. This virtuous process enables the unsupervised augmentation of multi-variant lexicons without requiring manual rule definition by experts. We ap- ply this multilingual methodology on Al- satian, a French regional language and provide (i) an intrinsic evaluation of the correctness of the obtained variants pairs, (ii) an extrinsic evaluation on a down- stream task: part-of-speech tagging. We show that in a low-resource scenario, collecting spelling variants for only 145 words can lead to (i) the generation of 876 additional variant pairs, (ii) a diminution of out-of-vocabulary words improving the tagging performance by 1 to 4%. 1 Natural Language Processing and Non-Standardized Languages Non-standardized languages present a great pro- ductivity of spelling variants for a given word. The absence of standardized spelling points up the ge- ographical and demographic variations that might exist and are otherwise smoothed down. This variability results in the coexistence of alternative written forms, hence in a large proportion of out- of-vocabulary words in the context of supervised machine learning. In what follows, we first present our approach to generate spelling variant pairs based on an ini- tial set of crowdsourced spelling variant pairs. This method is language independent and relies on resources that do not require expert knowledge, hence can easily be crowdsourced. Second, we exemplify the use of such a method to reduce the proportion of unknown words that undermines supervised algorithms in the context of non-standardized languages. 1.1 Working with Multi-Variant Linguistic Resources The question of variation in non-standardized lan- guages naturally arises starting when one be- gins the process of corpus building (or collec- tion). When dialectal and spelling variants over- lap, inter- and intra-dialectal variations can be hard, not to say impossible, to untangle. In the fol- lowing, we will design as “spelling variant” any variant due to either dialectal variation, spelling convention variation, or an accumulation of both. Although one might chose to work on corpora produced in a controlled environment, in which the spelling conventions and writers are carefully chosen, this setup is unlikely to produce satisfying results on real-life data. Producing linguistic resources, be it lexica, raw or annotated corpora, represents a cost that can- not be afforded for …

Content truncated to 500 words