A New Framework for Fast Automated Phonological Reconstruction Using Trimmed Alignments and Sound Correspondence Patterns

Abstract

Computational approaches in historical linguistics have been increasingly applied during the past decade and many new methods that implement parts of the traditional comparative method have been proposed. Despite these increased efforts, there are not many easy-to-use and fast approaches for the task of phonological reconstruction. Here we present a new framework that combines state-of-the-art techniques for automated sequence comparison with novel techniques for phonetic alignment analysis and sound correspondence pattern detection to allow for the supervised reconstruction of word forms in ancestral languages. We test the method on a new dataset covering six groups from three different language families. The results show that our method yields promising results while at the same time being not only fast but also easy to apply and expand.

The SIGTYP 2022 Shared Task on the Prediction of Cognate Reflexes

“…Overall, all systems do quite a good job at recovering unknown words from their cognate sets, specifically in those cases, where only a small part of the test data was retained for the evaluation process. Judging from our practical experience and independently published results on word prediction experiments (List et al, 2022b;Bodt and List, 2022), B-Cubed F-Scores higher than 0.7 and average edit distances of about 1 provide a good starting point for computer-assisted approaches and can already provide active help in various practical annotation tasks in historical linguistics. Thus, scholars working on the reconstruction of certain language families could use predicted proto-forms and later manually correct them, or field workers could use automatically predicted words when trying to elicit specific lexical items to search for cognate words that might have shifted their meanings.…”

“…Our baselines were taken from the reflex prediction framework by List et al (2022b). This framework consists of four major stages.…”

“…In stage (2), alignments are trimmed by merging all columns in the alignment in which the attested languages all show a gap with their preceding column. As a result, a word like Latin cenāre [k e: n a: r E] would be rendered as [k e: n a: r.E], when being aligned with Spanish cenar [T e n a R], since the final [E] in Latin corresponds to a gap in Spanish and could therefore not be predicted (see List et al 2022b for details on this procedure). In stage (3), alignments are enriched by coding for potentially conditioning context, which is added to the alignments in the form of additional rows.…”

“…(3) code context (4) train classifier Based on this general framework, we created two baselines, one primary baseline that uses the correspondence pattern recognition (CORPAR) method by List (2019a) as a classifier, and one extended baseline which predicts words with the help of a support vector machine (SVM, see List et al 2022b for details on both systems). From previous studies on supervised phonological reconstruction we know that the SVM variant of the framework outperforms the CORPAR classifier clearly, although differences are not extremely high (ibid.…”

“…In order to predict a word from several reflexes in different languages, the system first predicts individual target tensors of probabilities for each attested reflex and then averages them to produce the prediction. Team CEoT, represented by Tiago Tresoldi (Uppsala University), provided a workflow that predicts cognate reflexes based on phonetic alignments (Tresoldi, 2022), which is quite similar to the extended baselines of our shared task (List et al, 2022b). In contrast to our baseline approaches, their system EXTALIGN-RF skips the trimming procedure (stage 2), varies the techniques for alignment enrichment by taking preceding and following context into account (stage 3), and uses a random forests classifier rather than a support vector machine.…”