Ab Initio: Automatic Latin Proto-word Reconstruction

Alina Maria Ciobanu, Liviu P. DinuCOLING 2018

TL;DR

This paper presents a novel computational approach for automatically reconstructing Latin proto-words from their modern Romance language descendants. By leveraging sequence labeling and ensemble methods, the system achieves state-of-the-art performance while requiring less data than previous methods, making it valuable for historical linguistics research.

What Problem Does This Paper Address?

Historical linguistics aims to understand language evolution by reconstructing ancestral languages, known as proto-languages, from their modern descendants. A key task is reconstructing "proto-words" – the original forms of words in a proto-language. Traditionally, this is a manual, time-consuming process relying on the "comparative method," which involves identifying cognates (words with a common origin) and comparing their properties. While computational methods have emerged, they often require extensive data like phonetic transcriptions or large corpora, which are scarce for many endangered or extinct languages. This paper addresses the challenge of automating proto-word reconstruction for Latin, specifically focusing on developing a method that is accurate, efficient, and requires minimal input data, thus making it more accessible for historical linguists.

Methods

The proposed method operates in two main steps:

  1. Individual Language Proto-word Reconstruction: For each modern Romance language, the system uses a sequence labeling approach based on Conditional Random Fields (CRFs). Given pairs of modern words and their corresponding proto-words (from a training set), the CRF learns orthographic patterns of spelling changes. The input to the CRF is the modern word, and the output is a sequence of characters representing its proto-word. Features for the CRF include character n-grams from the input word. Before feeding words to the CRF, a pairwise sequence alignment is performed using the Needleman-Wunsch algorithm to establish correspondences between characters in the modern word and its proto-word.
  2. Ensemble Proto-word Reconstruction: To improve accuracy, the system combines the outputs from individual language models. Several ensemble methods are proposed, which aggregate the "n-best lists" (lists of the top-n most probable proto-word reconstructions) from each language. These methods fuse information based on:
  3. Rank: Averaging the ranks of potential proto-words across languages.
  4. Rank and Accuracy: Weighting language contributions by their training accuracy.
  5. Weight (Confidence Score): Averaging the confidence scores from the sequence labeling system.
  6. Weight and Accuracy: Combining confidence scores with language training accuracy.

The best-performing ensemble method uses a combination of rank and training accuracy.

Datasets

The paper utilizes three datasets of Romance languages with Latin ancestors:

  • Dataset 1: Proposed by Bouchard-Côté et al. (2007), containing 585 complete cognate sets in Spanish, Portuguese, and Italian, with their Latin ancestors. It's available in orthographic and phonetic (IPA) versions.
  • Dataset 2: Proposed by Reinheimer Ripeanu (2001), with 1,102 cognate sets in Spanish, Italian, Portuguese, French, and Romanian, and their Latin ancestors. This dataset includes incomplete cognate sets.
  • Dataset 3: Proposed by Ciobanu and Dinu (2014b), containing 3,218 complete cognate sets in Spanish, Italian, Portuguese, French, and Romanian, with their Latin ancestors.

All datasets were split into train, development, and test sets with a 3:1:1 ratio.

Key Insights and Ideas

  • Orthographic Regularities are Powerful: The study demonstrates that orthographic changes between modern Romance languages and Latin often reflect underlying sound correspondences, making it possible to reconstruct proto-words using only orthographic forms, without needing phonetic transcriptions.
  • Ensembling Boosts Performance: Combining predictions from individual language models significantly improves proto-word reconstruction accuracy compared to using any single language model.
  • Data Scarcity is a Major Hurdle: The paper highlights the advantage of its method in requiring less data, which is crucial for historical linguistics where resources are often limited.
  • Dataset Quality Matters: The performance of the system is significantly influenced by the quality and nature of the cognate sets in the dataset, with datasets built on established etymologies yielding better results than those based solely on edit-distance thresholds.

Key Results

  • The best-performing ensemble system achieved an average edit distance of 1.07 on Dataset 3, meaning reconstructed proto-words were, on average, just over one character different from the correct form.
  • The correct proto-word was found within the top 5 reconstructed words for 70% of cases on Dataset 3 (increasing to 74% for top 10).
  • On Dataset 1 (orthographic), the ensemble achieved an average edit distance of 2.31, outperforming the previous state-of-the-art method's 2.34.
  • The oracle classifier (a theoretical upper bound for ensemble performance) achieved an average edit distance of 1.76 on Dataset 1 (orthographic), indicating substantial room for improvement.
  • The system performs well even with incomplete cognate sets, a significant advantage for real-world linguistic data.

Limitations and Future Work

  • RNN Performance: While Recurrent Neural Networks (RNNs) were explored, they did not outperform the CRF-based system, likely due to their higher data requirements, which are problematic in historical linguistics. The authors suggest future work to improve RNN performance with limited data.
  • Error Analysis: The paper identifies specific error patterns, such as mistaken final vowels (e.g., 'um' instead of 'us') and issues with consonant doubling. For two-character differences, errors include 'f' instead of 'ph' and the suffix 'a' instead of 'us' (often resulting in feminine adjective forms instead of masculine).
  • Future Work:
    • Refine fusion methods for ensemble classifiers, inspired by oracle results.
  • Evaluate the method on datasets covering more languages, such as Swadesh lists or the Austronesian basic vocabulary database.
  • Investigate specific linguistic phenomena like rhotacism.
  • Further improve CRF system performance and explore RNN enhancements for low-data scenarios.

Summary updated

BibTeX
@inproceedings{ciobanu2018ab,
  author = {Alina Maria Ciobanu and Liviu P. Dinu},
  title = {{Ab Initio: Automatic Latin Proto-word Reconstruction}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure