Using feature vectors for automated phonological reconstruction and reflex prediction

Abstract

Introduction In recent years, automated phonological reconstruction and cognate reflex prediction have significantly evolved. Yet, these tasks remain a key challenge in computational historical linguistics due to data sparsity and the complexity of phonological systems. This paper introduces the Feature Vector Transformer (FeVeT), a new approach that addresses these limitations by representing phonological segments as structured feature vectors rather than discrete symbols. Methods The proposed model uses a Transformer-based encoder-decoder architecture to generate reflexes or proto-forms from cognate sets. Input characters are encoded as phonological feature vectors, incorporating language and positional information. The model is trained to predict both phonological features and IPA symbols simultaneously, enabling robust and interpretable outputs without requiring sequence alignment or trimming. Results FeVeT achieves state-of-the-art performance on both reflex prediction and proto-language reconstruction tasks across multiple datasets. Ablation studies confirm the importance of feature-based input representations, and pre-training on reflex prediction data improves generalization in the reconstruction task. Discussion The use of feature vectors enables FeVeT to capture fine-grained phonological patterns and generalize across languages with sparse training data. The model outperforms prior encoder-decoder and position-wise classification architectures in most settings. Challenges remain with irregular changes such as metathesis and reconstructing proto-forms in high test-proportion or zero-shot settings. Suggestions for future work include integrating phylogenetic information into language embeddings and refining pre-training strategies. Conclusion FeVeT represents a step forward in computational approaches to phonological reconstruction and reflex prediction. By incorporating phonological feature representations, it improves performance under data-sparse conditions, highlighting the potential of feature-based representations for future advances in the field.