Transfer of Models and Resources for Under-Resourced Languages Semantic Role Labeling

Identifying and labeling the semantic roles of words and phrases within a sentence is a crucial task in the field of natural language processing (NLP), known as semantic role labeling (SRL). Although SRL techniques have been well-developed for languages with high resources, under-resourced languages pose substantial challenges due to the scarcity of annotated data and language-specific tools. This paper investigates the potential of utilizing models and resources from languages with high resources to enhance SRL performance in under-resourced languages. The paper provides an overview of the SRL process, identifies the specific challenges faced by under-resourced languages, and proposes techniques for transferring models and resources from languages with high resources. We investigate a potential solution, which is based on a transfer learning approach that utilizes a pre-trained English SRL model. Our approach leverages a substantial parallel English-Amharic corpus to build the Amharic SRL dataset without the need for extensive manual annotation. Our model achieves an outstanding F1 score of 85.2% on the Amharic test set, showcasing its effectiveness in identifying and labeling semantic roles in Amharic sentences. Using transfer learning and extensive parallel corpora, our method overcomes data scarcity and proves the possibility of creating advanced language processing tools for under-resourced languages.