Abstract
Large-scale cross-lingual analysis of grammatical features like tense is often hindered by the lack of NLP tools for the majority of the world's languages. This paper presents SuperPivot, a novel computational method that leverages "superparallel" corpora (the Bible) to project linguistic markers from overt languages to non-overt ones. By utilizing Creole languages as high-precision "head pivots" and character n-gram alignments, we analyze tense across 1,169 languages. Our results demonstrate high accuracy (Mean Reciprocal Rank of 0.85) and provide a scalable framework for automated typological research in low-resource settings.
Introduction
Natural Language Processing (NLP) faces a significant "resource gap." While advanced tools exist for a few hundred languages, thousands remain "low-resource," making large-scale structural investigations difficult. Traditional typology is manual and slow, while previous computational methods required explicit marking across all target languages—a requirement that fails for languages expressing tense through context. We address this by introducing SuperPivot, which extracts linguistic information from parallel corpora without requiring pre-existing analysis tools.
The SuperPivot Methodology
SuperPivot relies on two core hypotheses: Overt Encoding (H1), which posits that any major linguistic distinction is clearly marked in at least some languages, and Projection (H2), which suggests these markers can be projected across a parallel corpus.
The workflow involves identifying a "head pivot" language with regular markers, typically a Creole (e.g., Seychelles Creole), and expanding this to a "pivot set" of the top 100 correlated markers across languages. These markers are then projected to 1,000+ languages using character n-grams (2-6 characters). This approach bypasses tokenization issues in languages like Chinese or Inuit.
Experimental Setup
We utilized the Parallel Bible Corpus (PBC), covering 1,169 languages across 7,958 verses. Evaluation on 10 diverse languages (e.g., Arabic, Chinese, English, Russian, Yoruba) used Mean Reciprocal Rank (MRR) to assess marker identification. A key innovation was the use of linear alignment, which assumes a marker's relative position in a verse is consistent across languages. When aggregated over 100 pivots, this method effectively smooths out noise from varying word orders.
Results and Typological Insights
The method achieved an average MRR of 0.85. Performance was near-perfect for English and German (MRR 1.00) but lower for Chinese (0.29) and Arabic (0.72). Chinese lacks grammaticalized tense, while Arabic's templatic morphology (internal vowel shifts) is less amenable to n-gram analysis.
Creole languages proved to be ideal pivots due to their regular grammatical structures and lack of historical irregularities. The resulting semantic maps allow researchers to visualize how different languages partition the conceptual space of time.
Challenges and Future Directions
While successful, SuperPivot faces challenges with polysynthetic and templatic languages. Additionally, the system occasionally conflates tense with aspect (e.g., in Slavic languages). Future improvements include clause-level alignment to reduce noise in complex sentences and the integration of past, present, and future markers into a unified temporal map.
Conclusion
SuperPivot represents a significant step toward automated, large-scale linguistic typology. By bridging the gap between high-resource and low-resource languages, it enables the study of grammatical structures across the full diversity of human language.
Summary updated