Less is More: Data-Efficient Complex Question Answering Over Knowledge Bases

Question answering is an effective method for obtaining information from knowledge bases (KB). In this paper, we propose NS-CQA, a data-efficient reinforcement learning framework complex question by using only modest number of training samples. Our consists neural generator and symbolic executor that, respectively, transforms natural-language into sequence primitive actions, executes them over the base to compute answer. We carefully formulate set actions that allows us not simplify our network design but also accelerate model convergence. To reduce search space, employ copy masking mechanisms in encoder-decoder architecture drastically decoder output vocabulary improve generalizability. equip with memory buffer stores high-reward promising programs. Besides, adaptive reward function. By comparing generated trial trials stored buffer, derive curriculum-guided bonus, i.e., proximity novelty. mitigate sparse problem, combine reshaping dense feedback. Also, encourage generate new avoid imitating spurious while making remember past data efficiency. NS-CQA evaluated on two datasets: CQA, recent large-scale dataset, WebQuestionsSP, multi-hop dataset. On both datasets, outperforms state-of-the-art models. Notably, performs well questions higher complexity, approximately 1% total