A Modular Dual Learning for Improving Question Answering and Generation Over Knowledge Graphs

Question answering (QA) and question generation (QG) over knowledge graphs (KGs) are inherently complementary, as they both entail the mapping of natural language to relations in KGs and the formation of reasoning paths. Existing studies show that KGQG can enhance the KGQA systems through data augmentation. An interesting discussion is whether KGQA can also benefit KGQG models. It can be postulated that the joint learning of KGQA and KGQG may result in mutual improvement, given that the two tasks are symmetrical in form. However, the effectiveness of their joint learning has not been rigorously investigated. This paper presents a modular dual learning framework for KGQA and KGQG, which fosters dual learning through an adaptively chosen and self-organized modular sharing mechanism to boost the performance of both tasks. The framework comprises three core modules: the KGQA module, the KGQG module, and the modular sharing network. Specifically, for a given input, the model initially generates a routing selector via a discrete latent variable in a mixture of experts fashion. Subsequently, the selector produces a routing that flexibly selects appropriate network layers from the shared network collection and organizes them into a network adapted to the current task. The shared neural layers are identical in structure and parameter size. In the construction of the new network, the shared layers can be positioned at various points along the sequential path and subsequently reused. As a result, these layers can contribute to the final output at distinct stages of processing. To enhance the current module by incorporating a dual task inductive bias, an assessment procedure is recommended to determine if the shared network should be included in the present computation. The proposed framework has been demonstrated to substantially boost the performance of KGQA and KGQG in experiments conducted on two widely-utilized benchmarks. The utilization of dual signals within the shared network has the potential to enhance the model's generalization capabilities. Furthermore, when the trained module is substituted with a randomized network, the modular shared network still demonstrates notable advancement of the new model with slight fine-tuning, which suggests the transferability of knowledge.