TL;DR
This paper presents a computational framework for predicting cognate words in extant languages by adapting a Bayesian phylogenetic inference approach. It tackles the challenge of reconstructing missing words in language families, offering a novel way to leverage sound correspondences and evolutionary relationships for linguistic reconstruction.
What Problem Does This Paper Address?
The paper addresses the challenge of reconstructing ancestral language forms, specifically focusing on predicting missing cognate words in extant languages. Traditional historical linguistics often relies on expert knowledge to identify cognates and reconstruct proto-forms. This work aims to automate this process by building a computational model that can infer a missing word from a cognate set within a language family. This is particularly relevant for tasks like the SIGTYP 2022 Shared Task, where the goal is to predict a masked word in a list of related languages. The core problem is to leverage the patterns of sound change and evolutionary relationships between languages to accurately predict unknown lexical items.
Methods
The core of the system is an adaptation of a Bayesian phylogenetic framework previously developed for reconstructing proto-language words. The workflow involves several key steps:
- Pairwise Hidden Markov Models (pHMMs): These models are trained to learn pairwise string alignments, capturing sound correspondences between words. The parameters are estimated using the Baum-Welch algorithm on training data.
- UPGMA Tree Construction: A preliminary phylogenetic tree is built using the UPGMA algorithm based on pairwise Levenshtein distances between languages. This tree serves as a guide for subsequent steps.
- Multiple Sequence Alignment (MSA): The T-Coffee algorithm is employed to perform multiple sequence alignments on cognate classes, using the trained pHMM parameters. This aligns words from multiple languages for a given concept.
- Bayesian Phylogenetic Inference: The aligned sequences are then used for Bayesian phylogenetic inference. This step models sound changes as a continuous-time Markov process on the language tree, inferring posterior distributions of phylogenies and mutation rates.
- Inferring Mutation Rates: Mutation rates for symbols within MSA columns are inferred, treating gaps as regular characters.
- Cognate Prediction (Missing Value Imputation): For the test data, the system aligns the non-masked entries of a cognate class. Then, using the inferred posterior distributions of trees and rates, it imputes the missing word column by column, selecting the symbol with the highest posterior probability.
Datasets
The paper utilizes datasets provided for the SIGTYP 2022 Shared Task on the Prediction of Cognate Reflexes. These datasets consist of parallel word lists from 20 language families. Each dataset contains between four and 19 related languages and between 500 and approximately 10,000 words. Words are pre-classified into cognate classes. The data is provided in IPA transcription and tokenized. For evaluation, each dataset was split into training and test sets, with test data proportions varying from 10% to 50%. For the test data, one word per row was masked, and the task was to predict this masked word.
Key Insights and Ideas
- Time-Reversible Model of Language Change: The paper highlights that the inference methods used are based on a time-reversible model of language change, which makes the approach suitable for both reconstructing ancestral forms and predicting extant forms.
- Leveraging Phylogenetic Structure: The system explicitly uses phylogenetic information (the language tree) to weigh the importance of cognates from different languages when making predictions. This acknowledges that languages more closely related evolutionarily might provide more relevant clues.
- Sound Correspondences Encapsulated in pHMMs: The trained pHMMs effectively capture regular sound correspondences, which are a fundamental aspect of cognate identification and reconstruction.
- Independence from Language-Specific Sound Laws: A notable aspect is that the prediction step does not directly use language-specific sound laws. Instead, it relies on the learned statistical patterns of sound correspondences from the data.
- No Syntagmatic Information Used: The current approach does not consider syntagmatic information (e.g., phonotactic rules or morphological similarity) when predicting words, focusing solely on sound correspondences and phylogenetic relationships.
Key Results
The paper describes the workflow and methodology for the SIGTYP 2022 Shared Task. While specific quantitative accuracy scores for the prediction task are not detailed in the provided text, the methodology is presented as an adaptation of a previously successful framework. The example provided illustrates the prediction of the French word "pifk" from a cognate set including Albanian "piski," English "fıf," German "fif," and Latin "piski." The system aims to produce the "maximum a posteriori" cognate prediction.
Limitations and Future Work
- Lack of Language-Specific Sound Laws: The system's reliance on general sound correspondences learned from data, rather than explicit language-specific sound laws, is noted as a limitation. This means it might miss nuances captured by known historical sound changes.
- No Syntagmatic Information: The absence of syntagmatic information (phonotactics, morphology) in the prediction process is acknowledged. This means the system doesn't check if the predicted word is phonologically or morphologically plausible within its language.
- Future Work: The authors suggest extending the system to incorporate language-specific sound correspondences and syntagmatic information in future research. This would likely lead to more accurate and linguistically informed predictions.
Summary updated