Bootstrapping a Neural Morphological Analyzer for St. Lawrence Island Yupik from a Finite-State Transducer

Abstract

Morphological analysis is a critical enabling technology for polysynthetic languages. We present a neural morphological analyzer for case-inflected nouns in St. Lawrence Island Yupik, an endangered polysythetic language in the Inuit-Yupik language family, treating morphological analysis as a recurrent neural sequence-to-sequence task. By utilizing an existing finite-state morphological analyzer to create training data, we improve analysis coverage on attested Yupik word types from approximately 75% for the existing finite-state analyzer to 100% for the neural analyzer. At the same time, we achieve a substantially higher level of accuracy on a held-out testing set, from 78.9% accuracy for the finite-state analyzer to 92.2% accuracy for our neural analyzer.

Bootstrapping Techniques for Polysynthetic Morphological Analysis

“…We observed an issue with syncretic ambiguity which complicates the evaluation process (also noted by Schwartz et al 2019;Moeller et al 2018).…”

“…Neural morphological analyzers can be developed from training data generated by an FST. These analyzers are more robust, handling variation, out-of-vocabulary morphs, and unseen tag combinations (Micher, 2017;Moeller et al, 2018;Schwartz et al, 2019). They provide 100% coverage, always providing a "best guess" analysis for any surface form.…”

“…They provide 100% coverage, always providing a "best guess" analysis for any surface form. Of course, FSTs can be modified to accommodate exceptions and OOV morphs, but this requires explicit modelling and usually does not achieve the robustness of neural analyzers (Schwartz et al, 2019). Anastasopoulos and Neubig (2019) found that they could augment their training set by hallucinating new stems, increasing accuracy on their test set by 10 percent.…”

“…Given our low resource setting, training a neural encoder-decoder model like those used in neural machine translation (NMT) is not possible without augmenting what resources we do have. Following the established template of recent work on neural morphological analysis for low resource polysynthetic languages (Micher, 2017;Moeller et al, 2018;Schwartz et al, 2019) we use the FST model to generate morphotactically valid pairs of surface and analyzed verbs.…”

“…The model setup is similar to the one described in (Schwartz et al, 2019). We use Mar-ianNMT: a fast, open-source toolkit which implements neural models for machine translation (Junczys-Dowmunt et al, 2018).…”