Question-Answering over DBpedia with Pretrained Auto-regressive Models

Project Summary

This project aims to tackle the challenge of generating formal SPARQL queries from natural language questions to streamline and simplify the extraction of information from DBpedia knowledge graph. This is an instrumental process in fostering intuitive, factual dialogues that are backed with open knowledge graphs such as DBpedia.

Among the exsiting approaches to this problem, i.e., classification, ranking, and translation, this proposal will primarily focus on the latter one. Building on top of the ongoing Neural SPARQL Machines project, this project aims to harness the significant performance improvement observed with the use of pre-trained large language models in translation tasks. However, our objective goes beyond merely refining the task of text-to-sparql translation; Our ambition is to deliver an intuitive interface to clear the path for the extensive adoption of open knowledge graphs across various sectors.

Mentors

Tommaso Soru, Anand Panchbhai, Saurav Yogen Joshi, Sanju Tiwari

Topics

Knowledge graphs, Transformers, Large Language Models, Question-Answering

---