Ethio-Semitic Proto-Language Reconstruction with In-Context Learning and LSTM Encode-Decode Model
No Thumbnail Available
Date
2024-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
As language evolve, it change and words obtain new meanings and lose old ones, making their reconstruction a critical area of study. Proto-EthioSemitic languages, in particular remain underexplored despite their cultural and historical significance. This research investigates Historical Language Reconstruction (HLR) for Proto-EthioSemitic languages in word level, focusing on two core objectives: cognate identification and proto-word reconstruction. A three-way dictionary was used to compile a dataset of 14,100 semantically related words from Amharic, Ge’ez, and Tigrinya. Linguists manually identified a golden data set with 74 cognate pairs from the Swadesh list concept translated into the three languages of interest and reconstructed proto-forms, while using automated methods (SCA and LexStat) extracted an additional 1,847 cognates from the dataset, significantly enhancing scale. Building on these results, synthetic proto-forms were generated using in-context learning with GPT-4o, based on its performance of achieving a reconstruction accuracy of 85% when evaluated against the golden data. Furthermore, an LSTM-based encodedecode model was trained on the generated data to predict proto-forms from cognates, achieving a prediction accuracy of 91% and an average edit distance of 0.21. This work establishes a foundation for reconstructing ancestral languages within the Afro-Semitic family by integrating linguistic expertise, automated cognate extraction tools, and state-of the-art large language models. The findings underscore the potential of interdisciplinary approaches in preserving and understanding linguistic heritage, with implications for future studies in historical linguistics and language preservation.
Description
Keywords
Cognates, Proto-word, In-context learning, GPT 4o, LSTM based encodedecode.