Abstract
Reconstructing ancestral "proto-languages" from modern descendants is a fundamental task in historical linguistics. While automated methods exist, they typically require extensive labeled datasets of cognate sets paired with gold-standard reconstructions. This paper proposes a semisupervised approach using a bidirectional neural architecture, the Daughters-to-Protoform-to-Daughters (DPD) BiReconstructor. By leveraging the linguistic principle of regularity, the model learns from limited labeled data and abundant unlabeled cognate sets, significantly outperforming supervised baselines.
Introduction
Historical linguists employ the "comparative method" to reconstruct extinct languages by analyzing cognates—words across different languages that share a common ancestor. For instance, comparing the word for "mother" in Romance languages allows for the reconstruction of the Latin mater.
While machine learning has begun to automate this process, existing models often depend on "full supervision," requiring massive datasets where every set of daughter words (reflexes) is already paired with a known reconstruction. In practice, confirmed reconstructions are scarce compared to the volume of available cognate sets. This research addresses the semisupervised reconstruction task, enabling models to learn from a small fraction of labeled data alongside a large corpus of unlabeled cognates.
Methodology: The DPD BiReconstructor
The authors introduce the Daughters-to-Protoform-to-Daughters (DPD) BiReconstructor, designed to reflect the linguistic "regularity principle," which posits that sound changes are systematic.
The architecture comprises three primary components:
- D2P (Daughter-to-Protoform): An encoder-decoder network that predicts the ancestral form from a set of daughter words.
- P2D (Protoform-to-Daughter): A reflex-prediction network that attempts to "evolve" the predicted protoform back into the original daughter words.
- Bridge Network: To maintain differentiability between the D2P decoder's hidden states and the P2D encoder, a dense "bridge" network is utilized. This allows gradients from reflex prediction to inform the reconstruction process.
To ensure the model does not learn to predict reflexes from incorrect reconstructions, a CRINGE loss is implemented. This loss penalizes the model if it successfully derives a correct daughter word from a faulty protoform, thereby incentivizing the D2P component toward higher precision.
Experimental Setup
The models were evaluated on two linguistic datasets:
- WikiHan: A Sinitic dataset for Middle Chinese reconstruction (8 daughter languages).
- Rom-phon: A phonetic Romance dataset for Latin reconstruction (5 daughter languages).
The researchers simulated semisupervised conditions by varying the percentage of labeled data (5%, 10%, 20%, and 30%).
Results and Discussion
The DPD-based strategies consistently outperformed supervised baselines, particularly in low-resource settings. At 10% labeling, the DPD variant achieved 40.50% accuracy on WikiHan (vs. 33.25% baseline) and 37.60% accuracy on Rom-phon (vs. 30.69% baseline).
Key Insights
- Linguistic Constraints as Supervision: The P2D task provides "weak supervision." A reconstruction is deemed likely incorrect if it cannot logically produce its descendants through regular sound changes.
- Phonetic Representation Learning: Clustering analysis revealed that DPD models learned linguistically meaningful phoneme embeddings, naturally grouping vowels, consonants, and specific categories like nasal vowels without explicit instruction.
- Error Proximity: Even when reconstructions were not exact, the models showed a significant decrease in Token Edit Distance (TED), indicating that predicted forms were phonetically closer to the ground truth than those of baseline models.
Conclusion and Future Work
While the DPD BiReconstructor marks a significant advancement, challenges remain. The P2D sub-network occasionally predicts correct reflexes from incorrect protoforms, suggesting a need for further refinement of the intermediate representation. Future research may incorporate phylogenetic tree structures to better model branching sound changes and explore even lower-resource scenarios to further reduce reliance on manual reconstructions.
Summary updated