Question Answering using Automatically Generated Semantic Networks – the case of Swahili Questions

Question Answering systems are important in natural language processing applications such as web search and information extraction. Various gold standard curated datasets exist for QA mostly in English. QAs are solved using syntactic, semantic or even learning systems. Low resource languages, such as Swahili, have not been easy to process due to few resources for training, testing and learning. By recognizing that Swahili generally follows subject-verb-object format, which is also the general structure of semantic networks, we use this fact to create a semantic network. Scarcity of public domain corpus has meant that we use Swahili frequently asked questions from public websites, and school texts for Swahili language. We create a semantic network and subject it to QA tasks. The results show good accuracy of about 80% for any typical set of one fact comprehension questions. Challenges such as scarcity of corpora and lack of gold standard test sets remain.