Abstract: 

Visually Impaired Persons (VIP) are susceptible to loss of information when observing visual data such as charts due to the inaccessibility of information. Charts need to be accessible due to their heavy use in the scientific community and industry. We explore ways to make them accessible by employing Knowledge Base Question Answering (KBQA) practices, a widely opted technique to deal with fact-backed question-answering. We propose a question-answering engine that answers the user's query about a data chart. To answer users' queries, we implement a pipeline consisting of two separate NER models that extract the subject and predicate from the query and feed them to their respective classifiers to perform the task of entity linking and predicate linking. Finally, the SPARQL engine binds the obtained subject and predicate in the query, executes it on the knowledge base - an instance of Resource Description Framework (RDF), and retrieves the answer. We evaluate our approach at three stages; Entity detection, Predicate detection, Answer Retrieval. We achieved exact match accuracy of 84% at both subject detection predicate detection, and the SPARQL queries retrieved answers with 74% accuracy on the knowledge base.

Evaluation Committee:

• Dr. Asim Karim (Advisor)
• Dr. Agha Ali Raza