Triple alignment-enhanced complex question answering over knowledge bases

Abstract

Program induction is a crucial paradigm for complex question answering over knowledge bases. In the existing learning framework, the predicted program is required to strictly align (word-by-word) with a gold program, which could cause over penalization for minor deviations. Meanwhile, due to the existence of synonyms, program induction question answering often fails to retrieve answer from the knowledge base because individual function argument cannot perfectly replicate the target argument. To address above misalignment problems, we propose a triple alignment-enhanced complex question answering (TACQA) method by incorporating global token alignment, function alignment, and argument alignment. First, apart from the classical global token alignment, the predicted functions are extracted and aligned separately with the gold functions, enabling efficient learning of implicit structural information related to the query framework of program from input questions. Second, an argument alignment is introduced to correct the ambiguous function arguments, which enhances the disambiguation processing efficiency of multi-argument by optimizing candidate pool construction and similarity calculation. The experiments on KQA Pro show that our method consistently outperforms the SOTA methods, demonstrating the effectiveness of triple alignment processing mechanism for simultaneously addressing function misalignment and argument ambiguity in program induction and further improving the model performance.