Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology

Garrett Nicolai, Kyle Gorman, Ryan CotterellACL 2020

TL;DR

The SIGMORPHON 2020 workshop proceedings showcase advancements in computational research for phonetics, phonology, and morphology. This year's workshop featured three shared tasks focusing on typologically diverse morphological inflection, multilingual grapheme-to-phoneme conversion, and unsupervised morphological paradigm completion, highlighting the growing need for robust models that can generalize across languages and linguistic phenomena.

What Problem Does This Paper Address?

This document comprises the proceedings of the 17th SIGMORPHON Workshop, an annual event dedicated to bridging computational linguistics with theoretical linguistics. The workshop aims to foster research in applying computational techniques to morphology, phonology, and phonetics. A key challenge in this field is developing models that can effectively handle the vast diversity of human languages, especially low-resource languages, and generalize well to unseen linguistic structures. This year's workshop specifically addressed this by focusing on typologically diverse languages and tasks that require generalization, such as unsupervised learning and cross-lingual transfer.

Methods

The SIGMORPHON 2020 workshop featured three shared tasks, each with its own set of methodologies and approaches presented by the participating teams.

  • Task 0: Typologically Diverse Morphological Inflection: This task focused on inflection generation for languages with diverse typological features, many of which are low-resource. Approaches likely involved various neural network architectures, transfer learning techniques, and methods for handling data scarcity. The emphasis was on a system's ability to generalize to new, typologically distinct languages.
  • Task 1: Multilingual Grapheme-to-Phoneme (G2P) Conversion: This task aimed to build G2P conversion systems for multiple languages and scripts, addressing the limitations of existing research that is often language-specific or relies on proprietary data. Teams likely employed sequence-to-sequence models, transformer architectures, and techniques for multilingual learning and data augmentation.
  • Task 2: Unsupervised Morphological Paradigm Completion: This task bridged the gap between supervised inflection generation and unsupervised morphological generation. The goal was to generate complete inflection tables from raw text and a lemma list without explicit supervision for the inflectional forms. Methods likely included unsupervised learning, clustering, and leveraging linguistic constraints.

Datasets

The SIGMORPHON 2020 workshop utilized several datasets for its shared tasks:

  • Task 0: Focused on typologically diverse languages, including many low-resource languages. Specific details on the exact languages and data sources are not explicitly listed in the preface but the task's goal implies the use of diverse inflectional data.
  • Task 1: Used words and their corresponding IPA pronunciations extracted from Wiktionary, covering 15 languages and scripts. This provided a publicly available resource for G2P research.
  • Task 2: Utilized raw text and lemma lists for unsupervised morphological paradigm completion. The specific corpora used are not detailed in the preface.

Key Insights and Ideas

  • The critical need for generalization: A recurring theme is the challenge of building models that can generalize beyond the languages and linguistic phenomena they were trained on, especially for low-resource languages.
  • The power of shared tasks: The SIGMORPHON shared tasks continue to be a driving force for innovation, pushing the boundaries of what's possible in computational morphology and phonology by providing standardized benchmarks and fostering competition.
  • Bridging theoretical and computational linguistics: The workshop highlights the ongoing and fruitful interaction between computational approaches and theoretical linguistic insights, demonstrating how each can inform and advance the other.
  • The rise of unsupervised and low-resource methods: The inclusion of tasks focused on unsupervised learning and low-resource settings underscores the growing importance of developing methods that do not rely on extensive labeled data.
  • Multilingualism as a key direction: The G2P task, in particular, emphasizes the move towards multilingual models that can handle multiple languages and scripts, reflecting the reality of linguistic diversity.

Key Results

The proceedings present a wide range of results from the participating teams across the three shared tasks. Specific quantitative results (e.g., accuracy scores, F1 measures) are detailed within the individual papers and are not summarized in the preface. However, the preface does mention:

  • Task 0: 8 out of 14 submissions were accepted, indicating a competitive review process.
  • Task 1: 9 teams submitted a total of 23 different systems.
  • Task 2: 3 teams submitted a total of 7 different systems.

The key outcome is the presentation of novel approaches and systems that advance the state-of-the-art in computational phonetics, phonology, and morphology, particularly in the areas of typological diversity, multilingualism, and unsupervised learning.

Limitations and Future Work

Limitations and future work are discussed within the individual papers presented in the proceedings. However, based on the preface, some general directions and challenges can be inferred:

  • Low-resource language challenges: Many of the languages involved in Task 0 are low-resource, presenting inherent difficulties in data availability and model training. Future work will likely focus on more effective techniques for few-shot and zero-shot learning in these contexts.
  • Cross-lingual transfer: The success of cross-lingual transfer remains a key area for improvement, especially for typologically distant languages.
  • Unsupervised learning effectiveness: The unsupervised nature of Task 2 highlights the ongoing research into how effectively models can learn linguistic structures without explicit supervision, and future work will aim to improve the robustness and accuracy of these methods.
  • Data scarcity and replicability: The G2P task aimed to address limitations in existing research due to proprietary data. Future work will likely continue to focus on creating and utilizing open datasets to ensure replicability and broader impact.

Summary updated

BibTeX
@inproceedings{nicolai2020proceedings,
  author = {Garrett Nicolai and Kyle Gorman and Ryan Cotterell},
  title = {{Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure