Abstract
Unsupervised morphological case marker extraction is a significant challenge, especially for low-resource languages. We present CaMEL (Case Marker Extraction without Labels), a novel method that leverages massively parallel corpora and noun phrase (NP) projection to identify case markers across 83 languages without human-annotated labels. By utilizing statistical filtering based on NP boundaries and cross-linguistic triangulation, CaMEL achieves an average F1 score of 45% on a diverse test set, demonstrating the viability of inducing grammatical structures from raw parallel text.
Introduction
In linguistics, "case" refers to the morphological markings on nouns (such as suffixes) that indicate their grammatical role—whether a word is the subject, the direct object, or a location. Understanding case systems is fundamental to processing a language's grammar. However, extracting these markers automatically is difficult due to the complexity of case systems, where syncretism (multiple cases sharing the same marker) and polysemy are common. Furthermore, most of the world's languages are "low-resource," lacking the massive labeled datasets required to train standard NLP models. CaMEL addresses these challenges by inducing morphological structures from scratch using only raw, parallel text.
Methodology
The CaMEL pipeline relies on the Parallel Bible Corpus (PBC) and a series of statistical filtering steps to isolate case markers from general vocabulary.
- NP Annotation & Projection: Since case is primarily a property of nouns, the search space is restricted to Noun Phrases (NPs). The system identifies NPs in English and projects those boundaries onto target languages using word-level alignments.
- Candidate Generation: The model extracts all word endings ($n$-grams) within these projected NPs as potential case markers.
- Inside/Outside Filtering: A true case marker should appear significantly more often inside an NP than outside of it. The authors use Fisher’s Exact Test to calculate an odds ratio and p-value for every candidate suffix. Suffixes statistically "anchored" to NPs are retained.
- Suffix Restriction: Based on cross-linguistic observations, the model specifically targets word endings, as case is most frequently marked via suffixes.
Key Insights: Cross-Linguistic Triangulation
A core innovation of CaMEL is the use of "deep cases"—universal semantic roles like Instrumental or Locative. While surface markers vary, parallel text allows the model to "triangulate" meanings. For example, a single syncretic marker in Latin (like -ibus) can be disambiguated by observing its alignment with multiple distinct markers in a language like Russian.
This distributional signal allows the recovery of grammatical functions by comparing the distribution of substrings across specific syntactic chunks in parallel corpora.
Evaluation and Results
The model was evaluated using a subset of 83 languages from the Parallel Bible Corpus, with UniMorph paradigms serving as a "silver standard" for ground truth.
- Performance: On a test set of 19 languages, CaMEL achieved an average F1 score of 45%, with a preference for precision (0.54) over recall (0.41).
- Ablation Study: The "Inside/Outside" filtering based on p-values was identified as the most critical component; without it, performance drops to near zero.
- Visualization: t-SNE plots confirmed that the model effectively clusters NPs based on their semantic roles, successfully "splitting" ambiguous markers into their constituent deep cases.
Limitations and Future Work
Despite its success, CaMEL has limitations. It currently only targets suffixes, missing languages that use prefixes or internal word changes (ablaut). Additionally, the system's quality is dependent on the accuracy of word-level alignments, and choosing the correct length for overlapping substring candidates remains a challenge. Future work will explore extending this method to other grammatical features like tense or aspect.
Summary updated