Abstract
This paper investigates the problem of determining grammatical gender for the nouns of a language starting with minimal resources: a very small list of seed nouns for which gender is known or via translingual projection of natural gender. We show that through a bootstrapping process that uses contextual clues from an unannotated corpus and morphological clues modeled with suffix tries, accurate gender predictions can be induced for five diverse test languages.
Introduction
Grammatical gender is a fundamental property of many languages, where nouns are categorized into classes (e.g., masculine, feminine, neuter) that govern agreement with associated words like articles, adjectives, and verbs. While some gender assignments follow biological sex (natural gender), many are arbitrary. For instance, "table" is feminine in French (la table) but masculine in German (der Tisch).
For Natural Language Processing (NLP) tasks such as machine translation or part-of-speech tagging, knowing a noun's gender is vital. However, comprehensive electronic dictionaries for gender are often unavailable for many languages. This research addresses the "cold start" problem: building a high-accuracy gender classifier for a new language using minimal annotated data.
Methodology
The proposed framework utilizes a robust bootstrapping approach that alternates between contextual clues and morphological patterns.
Seeding and Translingual Projection
The process begins with a small list of "seed" nouns. These can be manually selected or automatically projected from English by translating words with natural gender (e.g., "woman," "king," "aunt") into the target language. This translingual projection leverages the fact that natural gender is often consistent across cultures.
Contextual Bootstrapping
The system identifies words that frequently appear in the vicinity of seed nouns. For example, if a specific article consistently precedes masculine seeds, it is identified as a reliable masculine indicator. These indicators are then used to label new, unknown nouns within a large unannotated corpus.
Morphological Analysis via Trie Models
To handle nouns that lack clear contextual indicators, the system analyzes word endings using a Variable-length Suffix Interpolation model. By constructing a Trie (prefix tree) of word endings, the system can interpolate gender probabilities from specific long suffixes (e.g., -ation) to more general short ones (e.g., -n). This approach effectively manages the noise of short suffixes while capturing the strong signals of longer ones.
Iterative Learning
The system cycles between these two models, using context to identify new seeds and morphology to generalize across the entire vocabulary.
Experimental Setup and Results
The system was evaluated on five diverse Indo-European languages: Romanian, French, Spanish, Slovene, and Swedish.
The system demonstrated exceptional performance across all test languages:
- Romanian: Achieved 99.81% accuracy using context and 99.44% when combined with morphology.
- French & Spanish: Reached 98-99% accuracy in token-based evaluations.
- Swedish: Despite the challenge of a "common" vs. "neuter" system (which does not map to natural gender), the system achieved 98.84% accuracy using context clues.
- Web-Scale Induction: Experiments using Google search counts to find agreement patterns (e.g., comparing "noun + masculine_adj" vs. "noun + feminine_adj") yielded 100% accuracy for a Romanian noun sample.
Conclusion and Future Work
The core insight of this work is that grammatical gender is a system of constraints. By identifying words that "agree" with known seeds, the system can map the gender landscape of a language with minimal supervision.
Future work includes extending this research to non-Indo-European languages, such as Bantu languages, which feature more complex noun class systems. Additionally, addressing "gender-flexible" nouns and rare words that lack both context and recognizable suffixes remains an area for improvement.
Summary updated