Developing Amharic Question Answering Model Over Unstructured Data Source Using Deep Learning Approach

Low-resource languages like Amharic are lagging while many technological advancements are taking place in Question Answering (QA) research. From improving search engine results to enabling conversational agents or chatbots, QA systems have many applications. But previous works on Amharic QA haven’t made any leap towards employing current techniques, and these researches are limited. Although the volume of Amharic texts generated and written electronic documents are increasing, no modern Amharic QA models take advantage of these data sources. In this study, a factoid textual Amharic QA model using deep learning approach has been designed and developed. Given the three components of a textual QA, question analysis is modeled as a classification problem, document processing or passage retrieval is modeled as an answer selection problem, and the answer extraction is done using a NER tool. A word2vec model has been developed from 2.3 GB of free Amharic text. 10,785 questions and answers with five attributes have been prepared. In the question classification (QC) component, an experiment has been made with various deep neural network models like CNN, SimpleRNN, LSTM, GRU, Bi-LSTM, stacked Bi-LSTM, and attention-based Bi-LSTM. The attention-based Bi-LSTM has shown a better performance. A hybrid Bi-LSTM/CNN model has been used in the answer selection component, and cosine similarity metrics have been applied. The result of the answer selection component is a sentence housing an answer. A publicly available NER tool which is based on FLAIR has been used regarding the answer extraction component. At this stage, results from the previous components have been used to extract the final answer as a named entity. With a reasonably large Amharic QA dataset compared to datasets used in previous studies and without handcrafted rules, linguistic tools, and frameworks, this study’s end-to-end deep neural network models have outperformed previous Amharic QA systems.