A Modular Approach for Efficient Simple Question Answering Over Knowledge Base

In this work, we propose an approach for efficient question answering (QA) of simple queries over a knowledge base (KB), whereby a single triple consisting of (subject, predicate, object) is retrieved from a KB for a given natural language query. In fact, most recent state-of-the-art methods exploit complex end-to-end neural network approaches to achieve higher precision while making it difficult to perform detailed analysis of the performance and suffering from long execution time when training the networks. To this problem, we decompose the simple QA task in a three step-pipeline: entity detection, entity linking and relation prediction. More precisely, our proposed approach is quite simple but performs reasonably well compared to previous complex approaches. We introduce a novel index that relies on the relation type to filter out subject entities from the candidate list so that the object entity with the highest score becomes the answer to the question. Furthermore, due to its simplicity, our approach can significantly reduce the training time compared to other comparative approaches. The experiment on the SimpleQuestions data set finds that basic LSTMs, GRUs, and non-neural network techniques achieve reasonable performance while providing an opportunity to understand the problem structure.