Proceedings of the 20th SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Morphology

Garrett Nicolai, Eleanor Chodroff, Frederic Mailhot, Çağrı ÇöltekinACL 2023

TL;DR

This document presents the proceedings of the 20th SIGMORPHON workshop, a key event for researchers in computational morphology, phonology, and phonetics. It highlights advancements in neural and finite-state models for complex morphology, particularly in low-resource languages, and showcases results from shared tasks focused on morphological inflection and interlinear glossing.

What Problem Does This Paper Address?

The SIGMORPHON workshop addresses the ongoing challenge of computationally modeling the intricate systems of human language, specifically focusing on morphology, phonology, and phonetics. This includes developing robust methods for understanding and generating word forms, especially for languages with complex inflectional systems or limited digital resources. The workshop also explores the intersection of linguistic phenomena with human cognition, aiming to uncover universal patterns and biases that shape language structure and learning.

Methods

The papers presented at SIGMORPHON explore a variety of methods, including:

  • Neural Models: Advanced neural network architectures are employed for tasks like morphological inflection generation, cross-lingual transfer, and grapheme-to-phoneme conversion. This includes transformer-based models and sequence-to-sequence approaches.
  • Finite-State Models: Traditional but powerful finite-state transducers (FSTs) are utilized, especially for low-resource languages and for tasks requiring explicit control over linguistic rules.
  • Hybrid Approaches: Some work combines neural and finite-state methods, leveraging the strengths of both for tasks like grapheme-phoneme transduction.
  • Cognitive Modeling: Research investigates how human cognitive biases might influence linguistic universals, using experimental studies and analyzing patterns in language acquisition and use.
  • Shared Tasks: The workshop features shared tasks that drive research by providing standardized benchmarks and datasets for specific linguistic problems, such as morphological inflection generation and interlinear glossing.

Datasets

The workshop utilizes and contributes to several datasets:

  • UniMorph: A widely used dataset for morphological inflection, with ongoing efforts to revisit and amend its data for various languages (e.g., Central Kurdish).
  • Interlinear Glossed Text Corpora: Datasets for interlinear glossing are crucial for developing models that can automatically annotate linguistic data, particularly for language documentation.
  • Grapheme-to-Phoneme (G2P) Datasets: Various datasets are used for training and evaluating models that convert written text to its phonetic representation, often focusing on cross-lingual and low-resource scenarios.
  • Newly Introduced Benchmarks: The shared tasks themselves often define new benchmarks and evaluation methodologies for specific linguistic challenges, pushing the state-of-the-art.

Key Insights and Ideas

  • Cognitive Biases Shape Language: Recurrent cross-linguistic patterns in morphology appear to mirror universal features of human cognition, suggesting a deep link between how we learn and process language and the structures we find across languages.
  • Feature Minimization vs. Maximization in Phonology: The traditional approach of minimizing phonological features might be less empirically adequate and computationally desirable than feature maximization, which aligns better with intensional definitions of phonological patterns and natural classes.
  • Neural Models for Complex Morphology: Neural networks are proving increasingly capable of handling complex morphological phenomena, even in low-resource settings, though they are often compared against and sometimes augmented by traditional methods.
  • Importance of Language Documentation: The shared task on interlinear glossing highlights the critical need for automated tools to support language documentation, a vital but often resource-intensive endeavor.
  • Typological Diversity is Key: The workshop emphasizes the importance of studying a wide range of languages, especially those with complex morphology, to uncover universal linguistic principles and build more generalizable models.

Key Results

  • Shared Task Performance: The SIGMORPHON shared tasks showcase competitive results from various research groups on tasks like morphological inflection generation and interlinear glossing, demonstrating progress in these areas. Specific performance metrics (e.g., accuracy, F1 scores) are detailed within individual papers.
  • Cognitive Plausibility: Studies suggest that certain morphological generalization patterns, particularly in languages like Korean, align with cognitively plausible learning biases.
  • Cross-Lingual Transfer Success: Research demonstrates the effectiveness of cross-lingual transfer for morphological analysis, even for languages with limited resources, by leveraging models trained on related languages.

Limitations and Future Work

  • Low-Resource Language Challenges: While progress is being made, developing robust computational models for languages with scarce data remains a significant challenge. Future work will likely focus on more effective few-shot learning and transfer learning techniques.
  • Interpretability of Neural Models: Understanding why neural models succeed or fail in linguistic tasks, especially in phonology and morphology, is an ongoing area of research. Future work may involve developing more interpretable models or better analysis techniques.
  • Integration of Linguistic Knowledge: The interplay between data-driven neural approaches and knowledge-driven linguistic theories (like those based on finite-state methods or explicit feature structures) is a fertile ground for future research.
  • Broader Cognitive Links: Further research is needed to solidify the connections between cognitive biases and observed linguistic universals, potentially involving more extensive experimental designs and cross-linguistic comparisons.

Summary updated

BibTeX
@inproceedings{nicolai2023proceedings,
  author = {Garrett Nicolai and Eleanor Chodroff and Frederic Mailhot and Çağrı Çöltekin},
  title = {{Proceedings of the 20th SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Morphology}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure