Bayesian-error-informed contrastive learning for knowledge-based question answering systems

Abstract

The Knowledge-base Question Answering (KBQA) system aims to answer a question based on a knowledge base (KB). However, incomplete knowledge bases (KBs) limit the performance of KBQA systems. To address this issue, we propose a contrastive regularization method that considers two modules to tackle this problem: knowledge expansion and a contrastive loss function, Bayesian-error-informed Contrastive Learning (BeCoL). These modules leverage latent knowledge from context KBs and their associated question–answer pairs to generate more such pairs. Additionally, we use these question–answer pairs for informative representation learning, which makes hard positive pairs attract and hard negative pairs separate. This approach will enhance the ability of the system to distinguish the pairs better, ultimately improving the systems performance. We evaluate our proposed approach on the WebQuestionSP (WebQSP), ComplexWebQuestions (CompWebQ), and GrailQA datasets. The results indicate that our approach outperforms existing methods across different KB settings in the WebQSP dataset at 10%, 30%, 50%, and 100% with Hits@1 scores of 43.8, 49.7, 61.3, and 73.7 respectively, and with F₁-scores of 28.2, 32.5, 44.3, and 61.1 respectively. Similarly, we achieved Hits@1 score of 52.7 and F₁-score of 44.2 on the CompWebQ dataset with 100% KB setting. For the GrailQA dataset under the 100% KB setting, our method attained an Exact Match (EM) score of 67.5 and an F${}_1$-score of 76.4. The findings demonstrate the proposed methods capacity to address low-resource settings and significantly improve the performance of KBQA systems. The code is available at https://github.com/ysudarshan-collab/BeCoL.