Complex Question Decomposition Based on Causal Reinforcement Learning

Complex question decomposition is an important research topic in the field of natural language processing (NLP). It refers to the decomposition of a compound question containing multiple ontologies and classes into a simple question containing only a single attribute or entity. Most previous studies focus on how to generate simple questions using a single attribute or entity but pay little attention to the generation order of simple questions, which may lead to an inaccurate decomposition or longer execution time. In this study, we propose a new method based on causal reinforcement learning, which combines the advantages of the current optimal performance reinforcement learning method and the causal inference method. Compared with previous methods, causal reinforcement learning can find the generation order of sub-questions more accurately, so as to better decompose complex questions. In particular, the prior knowledge is extracted using the counterfactual method in causal reasoning and is integrated into the policy network of the reinforcement learning model, and the reward rules of reinforcement learning are designed from the perspective of symmetry (positive reward and negative punishment), thus the intelligent body is guided to choose the sub-question with a greater benefit and less risk of decomposing. We compare the proposed method with the baseline method on three datasets. The experimental results show that the performance of our method is improved by 5–10% compared with the baseline method on Hits@n (n = 1, 3, 10), which proves the effectiveness of our proposed method.