A Computational Model for the Linguistic Notion of Morphological Paradigm

Miikka Silfverberg, Ling Liu, Mans HuldenCOLING 2018

TL;DR

This paper introduces a computational model that better captures linguists' understanding of morphological paradigms, moving beyond simple longest common subsequence (LCS) methods. It also presents a novel approach to learning complete inflectional paradigms from incomplete data, addressing the "paradigm cell filling problem" where not all word forms for a given lexeme are known.

What Problem Does This Paper Address?

Languages often have complex systems of word formation, where a single word can take many different forms depending on its grammatical role (e.g., tense, number, case). Linguists call these sets of related word forms a "morphological paradigm." While humans can intuitively generalize these patterns, computational models have struggled to accurately capture this. Existing methods, like using the longest common subsequence (LCS) of known word forms, often over-generate distinct paradigms for what linguists consider the same underlying pattern (e.g., due to vowel harmony). Furthermore, real-world data is often incomplete, meaning we might only have a few inflected forms for a word. This "paradigm cell filling problem" (PCFP) is a significant challenge for building robust morphological models.

Methods

The paper builds upon previous work using LCS to extract initial paradigm structures. The core contributions are:

  • Generalized Paradigms: They introduce a "phonology-free" method to group similar LCS-extracted paradigms that exhibit regular alternations (like vowel harmony). This is achieved by abstracting away varying parts of the paradigm with variables and then identifying patterns in these variable substitutions. This allows for a more linguistically plausible reduction in the number of distinct paradigms.
  • Learning Full Paradigms from Partial Data: To address the PCFP, they propose an algorithm that takes partial inflection tables (tables with missing forms) and imputes the missing forms. This involves:
  • Initial LCS-based Paradigm Extraction: Similar to the generalized paradigm method.
  • Stem Sampling: A crucial step to counteract incorrectly identified stems that can arise from sparse data. This involves sampling potential stem variations and using statistical evidence across all paradigms to find the most likely stem.
  • Imputation Methods: They evaluate several methods for filling in the missing forms, including majority voting, priority voting (a refinement of majority voting), and matrix completion.

Datasets

The authors use two primary datasets:

  1. Durrett and DeNero (2013) dataset: This dataset, scraped from English Wiktionary, includes Finnish, German, and Spanish data for nouns, adjectives, and verbs. It's used for both paradigm generalization and learning full paradigms from partial data.
  2. CoNLL-SIGMORPHON 2017 shared task data: This dataset, also from Wiktionary, covers 33 languages and is used exclusively for the paradigm generalization task due to its focus on full inflectional tables.

Key Insights and Ideas

  • Linguistic Intuition vs. LCS: The LCS method, while useful, fails to capture linguistically motivated generalizations like vowel harmony, leading to an over-segmentation of paradigms.
  • "Phonology-Free" Generalization: It's possible to group similar paradigms based on formal pattern matching of variable substitutions, without needing explicit phonological rules. This aligns better with how linguists group words.
  • Stem Sampling is Crucial for Sparse Data: When dealing with incomplete inflection tables, the initial identification of the word stem can be highly unreliable. A sampling approach that explores different stem possibilities and uses global evidence is essential for accurate paradigm completion.
  • Voting Methods are Surprisingly Effective: For learning full paradigms from partial data, simple majority and priority voting methods perform remarkably well, often outperforming more complex matrix completion techniques.

Key Results

  • Paradigm Generalization: The proposed method significantly reduces the number of distinct paradigms, bringing them closer to linguist counts. For Finnish, their generalized paradigm counts for verbs and nouns closely match linguistic analyses.
  • Full Paradigm Learning:
    • The "sampling + priority voting" method consistently outperforms other methods, especially when the number of input forms is small.
  • For Spanish verbs with only six input forms, this method reconstructs 83.64% of missing forms and 53.91% of complete paradigms.
  • For German nouns, with even fewer forms per paradigm, they achieve over 95% reconstruction of missing forms and over 92% of complete paradigms.
  • Matrix completion performed poorly across most datasets, except for German nouns.

Limitations and Future Work

  • Data Requirements: The paradigm generalization method requires full inflection tables for a given part-of-speech. This limits its direct application to languages or parts-of-speech where such complete tables are not available.
  • Neural Models: While neural models are powerful for inflection generation, their "black box" nature makes it difficult to extract interpretable linguistic generalizations. The authors suggest that future work could explore neural methods for full paradigm learning, but acknowledge the trade-off in interpretability.
  • Stem-Neural Model Integration: The current method automatically splits words into stems and inflectional material, which is a useful output for linguistic descriptions. Integrating this with neural models could be a future direction.
  • Further Exploration of Stem Sampling: The paper mentions that stem sampling provides consistent improvement, and further research could explore its impact more deeply, especially as the number of input forms increases.

Summary updated

BibTeX
@inproceedings{silfverberg2018computational,
  author = {Miikka Silfverberg and Ling Liu and Mans Hulden},
  title = {{A Computational Model for the Linguistic Notion of Morphological Paradigm}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure