Abstract
We present a system that automatically groups verb stems into inflection classes, performing a case study of Abui verbs. Starting from a relatively small number of fully glossed Abui sentences, we train a morphological precision grammar and use it to automatically analyze and gloss words from the unglossed portion of our corpus. Then we group stems into classes based on their cooccurrence patterns with several prefix series of interest. We compare our results to a curated collection of elicited examples and illustrate how our approach can be useful for field linguists as it can help them refine their analysis by accounting for more patterns in the data.
Introduction
Field linguistics is a race against time. Documenting endangered languages involves collecting thousands of recorded sentences, but the primary bottleneck is the manual process of creating Interlinear Glossed Text (IGT)—the word-by-word labeling of grammatical features.
A major challenge in this process is identifying word classes. In many languages, words belong to specific sub-classes based on which grammatical markers, such as prefixes, they can accept. In Abui, a language of Indonesia, verbs take different prefixes depending on the "undergoer" role (e.g., Patient vs. Recipient). Manually tracking which of the hundreds of verbs can take which of the many prefix combinations is a monumental task prone to human error and oversight.
Methods
The researchers leverage the AGGREGATION project pipeline to automate this distributional analysis. Their approach follows a "bootstrap" logic designed to work with the small, sparse datasets typical of endangered language documentation.
The pipeline involves several key steps:
- Inference: The MOM (Matrix-ODIN Morphology) tool analyzes a small set of manually glossed sentences to infer "morphotactics"—the rules governing which prefixes attach to which stems.
- Grammar Building: These rules are converted into a machine-readable precision grammar using the LinGO Grammar Matrix.
- Expansion: Using this grammar and the ACE parsing engine, the system automatically analyzes and glosses the remaining unglossed portion of the Abui corpus.
- Classification: Verbs are grouped into 16 possible inflectional classes based on a 4-bit binary code (e.g.,
1111indicates compatibility with all four tested prefix conditions).
Datasets
The study utilized two primary datasets:
- The Abui Corpus: A collection of approximately 18,000 sentences. At the start of the study, only ~4,600 were manually glossed.
- The Curated Set: A "gold standard" of 337 verb stems with expert-elicited judgments from native speakers regarding prefix compatibility, used to evaluate the system's accuracy.
Key Insights and Results
The system successfully provided morphological analyses for 4,642 previously unglossed words, significantly expanding the usable data for the language. When comparing the corpus to the expert "gold standard," adding the autoglossed data increased the number of matches from 248 to 287.
A significant finding was the tension between corpus data and speaker intuition. While elicitation (asking a speaker) shows what is theoretically possible, a corpus shows what is actually said. The system identified "mismatches" where a verb appeared with a prefix in the corpus that the expert hadn't recorded. This serves as a "discovery" tool for linguists; for example, the system identified a group of verbs (Class 0001) that were missing from the initial classification, which were later found to share specific semantic features related to "experiencer" arguments.
Conclusion
While quantitative performance metrics like F1 scores remained relatively low (improving from 0.10 to 0.19), the authors argue that the qualitative value—finding exceptions and new patterns—is more important for field linguists than raw accuracy. Future work will focus on integrating a phonological analyzer to handle spelling variations and extending the system to look at secondary prefix positions for even more granular classification.
Summary updated