TL;DR
This paper introduces VORM, a novel unsupervised morphological segmentation system that leverages parallel translation data to infer morphological transformations. VORM achieves competitive results on canonical segmentation tasks and introduces a new, typologically diverse dataset of 37 languages, pushing the boundaries of unsupervised morphological analysis beyond common European languages.
What Problem Does This Paper Address?
Morphological segmentation, the task of breaking down words into their constituent morphemes (meaningful units), is crucial for many natural language processing tasks, especially for morphologically rich languages. While supervised methods achieve high accuracy, they are limited to languages with extensive annotated data. Unsupervised methods, which learn from raw text, offer a path to segmenting words in a much wider range of languages. However, existing unsupervised methods often struggle with complex morphological processes like infixation and reduplication, and are typically evaluated on a narrow set of languages, primarily from Eurasia. This paper aims to bridge this gap by developing an unsupervised system that can handle a broader range of morphological phenomena and is evaluated on a more linguistically diverse dataset.
Methods
VORM is a three-step heuristic system:
- Determining Morphological Families: This step uses parallel translation data (bitexts) and word alignments to identify groups of related words. Instead of relying solely on distributional semantics, VORM leverages the idea that morphologically related words in one language will often translate to related words in another. This helps constrain the search space and avoid oversegmentation.
- Learning Productive Transformations: Inspired by previous work like Morsel, this step uses the identified morphological families to learn "productive" transformations (e.g., suffixes, infixes, reduplication). It employs a best-first heuristic to prioritize transformations that model the most word pairs within these families.
- Propagation to the Full Word List: The learned transformations are then applied to a larger wordlist, extending the segmentation beyond the initial bitext. This step uses a chain-based approach, where derivations are built by composing transformations, and a further constraint mechanism to ensure the coherence of the learned structures.
A key novelty is the use of translation equivalency to constrain the hypothesis space, allowing VORM to infer a wider array of morphological processes, including infixation and reduplication, which are often missed by other unsupervised methods.
Datasets
- Morphochallenge 2010 (MC10): Used for evaluation, this dataset includes gold data for canonical and surface segmentation for English, Finnish, Turkish, and German.
- SIGMORPHON 2022 (SGM22): Used for evaluation, this dataset focuses on surface segmentation for eight languages.
- DORECO-MORPH: A novel dataset introduced in this paper. It is derived from the DoReCo corpus, containing interlinearly-glossed fieldwork data for 37 typologically and areally diverse languages. This dataset allows for the generation of both supervised and unsupervised training and evaluation data, with a reproducible train/development/test split.
Key Insights and Ideas
- Translation Data as a Morphological Signal: The paper demonstrates that parallel translation data is a powerful signal for unsupervised morphological segmentation, effectively constraining the search space and improving the identification of morphological families.
- Canonical vs. Surface Segmentation are Different Tasks: The results highlight a significant performance gap between VORM on canonical segmentation (where it excels) and surface segmentation (where it is outperformed by models like Morfessor). This suggests that models optimized for one may not perform as well on the other, and that canonical segmentation is a more linguistically insightful metric for many purposes.
- Diversity is Key: The introduction and use of the DORECO-MORPH dataset underscore the need for evaluating unsupervised morphological models on a wider range of languages to capture diverse morphological phenomena like reduplication, which are underrepresented in existing benchmarks.
- Reduplication Handling: VORM shows a notable ability to correctly analyze reduplication, a process often missed by other unsupervised methods, as demonstrated by examples in Vera'a.
Key Results
- Competitive Canonical Segmentation: VORM achieves state-of-the-art or competitive results on canonical segmentation tasks, particularly outperforming other unsupervised models on the MC10 dataset for Finnish, German, and Turkish. On the DORECO-MORPH dataset, VORM achieves an average EMMA-2 score within 2% of the supervised Chipmunk model.
- Surface Segmentation Limitations: VORM is significantly outperformed by models like Morfessor on surface segmentation metrics (BPR), likely due to its focus on canonical forms and potential for undersegmentation.
- Ablation Study Insights: The ablation experiments reveal that while the compounding component (-CMPD) has a negligible or sometimes negative effect, and Step 1 (morphological families) can be redundant, omitting Step 3 (propagation to the full word list) consistently leads to a substantial drop in performance.
Limitations and Future Work
- Surface Segmentation Performance: VORM's performance on surface segmentation is a noted limitation, with lower recall scores compared to other models.
- Impact of Compounding and Step 1: The ablation studies suggest that the compounding component and the initial step of identifying morphological families might not always be beneficial, possibly due to their interaction with other components or the nature of the data.
- Future Work: The authors suggest further exploration of the DORECO-MORPH dataset by breaking it down into linguistically interesting subsets (e.g., languages with reduplication, assimilation processes) to identify more specific modeling challenges and gain a clearer understanding of unsupervised learning in morphology. They also propose investigating more specific modeling challenges through detailed analysis of model performance on different linguistic phenomena.
Summary updated