Low-resource languages often suffer from a lack of available data and resources, which makes it difficult to develop effective natural language processing (NLP) systems for these languages. However, with the increasing availability of graph-structured data, there is an opportunity to leverage this structure to generate text in low-resource languages.
Previous work on graph-to-text generation has achieved outstanding performance using large language models . However, the ability of these models to generate text in low-resource languages has not been thoroughly studied. The objective of this thesis is to explore the use of graph structure to generate text in low-resource languages. Specifically, the thesis will focus on developing a novel text generation approach that leverages graph-structured data, such as knowledge graphs, to generate high-quality text in low-resource languages.
• Solid programming skills (e.g. Python).
• Strong interest in natural language processing and machine learning.
• Experience in pre-trained language models or HuggingFace library is a plus.
Ausschreibung: Download (pdf)