Model for Semantic Network Generation from Low Resource Languages as Applied to Question Answering – Case of Swahili
Barack Wanjawa, Lawrence Muchemi · 2021 IST-Africa Conference (IST-Africa) · 2021
Though it is well recognized that Question Answering systems play an important role in natural language processing applications, such as web search and information extraction, their adoption for low resource languages has been low. This is mainly due to lack of language resources such as machine learning tools and training datasets. This research contributes to language processing initiatives by providing a model for semantic network generation for such languages. The model can guide on how to process natural language text and give it structure. The structured language can then be used for downstream tasks such as Question Answering. Proof of concept experiment is done using the Swahili part of TyDi QA, a publicly available QA dataset. The results show that the model can be used to generate semantic networks that can be applicable to Question Answering, with accuracy of 64.8% based on the sampled data. Lack of diverse and task specific curated corpora however remain a challenge in developing tools for low resource languages.