Abstract
We describe a supervised approach to predicting the set of all inflected forms of a lexical item. Our system automatically acquires the orthographic transformation rules of morphological paradigms from labeled examples, and then learns the contexts in which those transformations apply using a discriminative sequence model. Because our approach is completely data-driven and the model is trained on examples extracted from Wiktionary, our method can extend to new languages without change. Our end-to-end system is able to predict complete paradigms with 86.1% accuracy and individual inflected forms with 94.9% accuracy, averaged across three languages and two parts of speech.
Introduction
In morphologically rich languages like German, Spanish, or Finnish, a single "base" word can have dozens of inflected forms depending on tense, person, number, and case. Traditionally, capturing these patterns required linguists to laboriously hand-code complex rule systems. While unsupervised methods exist, they often struggle with accuracy.
This paper presents a supervised, data-driven system that learns to predict the entire set of inflected forms (a morphological paradigm) for a given word. By leveraging structured data from Wiktionary and jointly modeling all inflected forms together, the system achieves state-of-the-art accuracy across multiple languages without requiring hand-engineered rules.
Methodology
The authors propose a two-stage supervised approach to morphological prediction:
1. Rule Extraction
The system first learns an inventory of transformation rules by aligning base forms to their inflected counterparts in a training set. They use a modified edit-distance algorithm to identify "changed spans" (prefixes, suffixes, or internal stem changes). Crucially, they merge these spans across all forms in a paradigm to identify consistent points of transformation.
2. Discriminative Prediction
To inflect a new word, the system uses a semi-Markov Conditional Random Field (CRF). This model treats the base word as a sequence of characters and decides which rules to apply to which segments. It uses local orthographic features, such as character n-grams, to weigh different possible rules.
A key innovation is joint prediction. Instead of predicting each inflected form independently, the model predicts the entire table at once. This forces the system to adopt a single, consistent analysis of the base form across all its variants, which significantly reduces errors.
Datasets and Evaluation
The primary data source is Wiktionary. The authors built a scraper to extract complete inflection tables for German (nouns and verbs), Spanish (verbs), and Finnish (nouns and verbs). They also utilized CELEX, a standard lexical database, to evaluate German verb inflection against previous research.
Key Results
- High Accuracy: The system achieved an average of 86.1% accuracy on complete paradigms and 94.9% on individual inflected forms across three languages.
- Outperforming Baselines: On the CELEX German verb task, the model outperformed the previous state-of-the-art semi-supervised model, despite using a simpler, more scalable exact inference method.
- Language Independence: The same system worked effectively across German, Spanish, and Finnish with no language-specific tuning, proving its generalizability.
Conclusion and Future Work
The paper demonstrates that crowd-sourced, semi-structured data like Wiktionary is a high-quality substitute for expert-curated linguistic databases. Even without deep linguistic knowledge, a model looking only at character patterns can learn highly sophisticated morphological transformations.
However, the model currently relies purely on orthography and fails on "hidden" features, such as the gender of German nouns. Future work could incorporate unlabeled raw text to handle rarer words and develop more sophisticated alignment for complex cases like German separable-prefix verbs or Spanish reflexive verbs.
Summary updated