Principal Parts Detection for Computational Morphology: Task, Models and Benchmark

Principal parts-defined as the minimal set of cells from which all other forms within a lexeme s inflectional paradigm can be deduced-are an important concept in theoretical morphology. This concept, which outlines the minimal memorization needed for a perfect inflector, has been largely overlooked in computational morphology despite impressive advances in the field over the past decade. In this work, we formalize PRINCIPAL PARTS DETECTION as a computational task under the static scheme assumption, identifying a single set of cells as principal parts uniformly applicable across lexemes within a syntactic category. We construct a multilingual dataset of verbal inflection tables with gold principal parts annotations for ten typologically diverse languages. We evaluate several computational models for PRINCIPAL PARTS DETECTION, each implementing the same three-stage framework: characterizing relations between each pair of cells, clustering the resulting vector representations, and selecting a representative cell from each cluster as a predicted principal part. Our best-performing model, combining Edit Scripts between inflections with Hierarchical K-Means clustering, achieves an average F1 score of 55.05%, significantly outperforming a random baseline of 21.20%. While these results demonstrate initial success, further research is needed to advance PRINCIPAL PARTS DETECTION, which could optimize inputs for morphological inflection models and encourage exploration into the theoretical and practical significance of compact morphological representations.