KU-CST at the SIGMORPHON 2020 Task 2 on Unsupervised Morphological Paradigm Completion

Manex Agirrezabal, Jürgen WedekindACL 2020

TL;DR

This paper presents a novel unsupervised approach to discovering morphological paradigms (word families) from raw text and a list of lemmas. While not outperforming the official baseline, their method offers a promising, neural-network-free pipeline for uncovering linguistic structures, particularly relevant for low-resource languages.

What Problem Does This Paper Address?

Understanding how words change their form to convey different grammatical meanings (like tense, number, or case) is a fundamental challenge in linguistics and natural language processing. This is known as morphological analysis and generation. Traditionally, this has been tackled with rule-based systems. More recently, statistical and neural methods have emerged. However, the SIGMORPHON 2020 Task 2 presented a particularly difficult challenge: unsupervised morphological paradigm completion. This means the system had to discover these word families without any pre-existing knowledge of the language's morphology, relying solely on a raw text corpus and a list of base word forms (lemmas). This is crucial for understanding languages with limited digital resources, where explicit morphological dictionaries are unavailable.

Methods

The authors propose a four-step pipeline that makes a strong assumption: all inflected forms can be represented as a STEM + SUFFIX.

  1. Splitting: They systematically split every lemma and every word in the corpus into all possible stem-suffix combinations.
  2. Identifying Regular Inflections: For each potential stem, they identify the set of possible suffixes that appear in the corpus. They then use this to estimate the "morphological richness" of a lemma (how many distinct forms it can take). To handle semi-regular verbs, they compare pairs of lemmas with similar stems and identify common suffixes.
  3. Clustering Suffixes: The collected suffixes are then clustered using K-Means. The idea is that similar suffixes (even with minor variations due to morphophonological changes) will group together, representing distinct grammatical functions.
  4. Generating Paradigms: Finally, they combine the most frequent stem for each lemma with the clustered suffixes. A character-level language model is used to select the most probable stem-suffix combination for each potential word form.

A key innovation is the "Flexible model," which augments the initial lemma list by training a simple verb classifier. This allows the pipeline to discover more suffixes and potentially improve recall.

Datasets

The primary data source is the Bible, used as the raw text corpus. A list of verbal lemmas was also provided. The languages used for development included Maltese, Persian, Portuguese, Russian, and Swedish. Testing was performed on Basque, Bulgarian, English, Finnish, German, Kannada, Navajo, Spanish, and Turkish.

Key Insights and Ideas

  • The STEM + SUFFIX Assumption: While a simplification, this core assumption allows for a systematic approach to splitting words and identifying potential morphological units.
  • Leveraging Corpus Frequency for Stem Identification: The method relies on the frequency of stem-suffix pairings in the corpus to determine the most likely stem for a given lemma.
  • K-Means for Suffix Grouping: Using K-Means to cluster suffixes is a clever way to group morphologically similar endings, even when they aren't identical. This helps in identifying distinct grammatical roles.
  • The "Flexible Model" for Lemma Expansion: Augmenting the initial lemma list with predicted verbs is a practical way to increase the coverage of the system and discover more potential suffixes.

Key Results

The authors' models (both Non-flexible and Flexible) did not outperform the official baseline in any of the tested languages. However, the results show some promise:

  • For development languages, Portuguese and Swedish were best captured by the Non-flexible model.
  • For test languages, Spanish and English showed the best performance with the Non-flexible model.
  • The Flexible model, while potentially having better recall, produced results that still required significant filtering.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • The STEM + SUFFIX Assumption: This assumption is too simplistic for languages like German, which utilize prefixes (e.g., ge- in gespielt). Future work could incorporate prefix handling.
  • Estimating Morphological Richness: The current method for estimating rm (the number of possible inflections) is an approximation. More robust methods, perhaps involving clustering of lemma triples, could be explored.
  • Neural Network Integration: The current pipeline is entirely rule-based and does not use neural networks. The authors believe incorporating Sequence-to-Sequence models would be highly beneficial for learning complex morphophonological changes.
  • Handling Compound Verbs: The model could be improved by better handling of compound verbs (e.g., X+V).
  • Orthographic Inconsistencies: Differences between lemma orthography and corpus orthography (as seen in Basque) pose a challenge and might require preprocessing.

Summary updated

BibTeX
@inproceedings{agirrezabal2020ku,
  author = {Manex Agirrezabal and Jürgen Wedekind},
  title = {{KU-CST at the SIGMORPHON 2020 Task 2 on Unsupervised Morphological Paradigm Completion}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure