Adapting Contrastive Learning with Feature Fusion for Complex Question Answering

Answering natural language questions on knowledge bases (KBs) has been a research hotspot. Great efforts have been made on complex question answering, due to urgent demands. Given a complex question, existing approaches mostly transform it into query graphs via semantic parsing and pick the best one for result identification. However, these approaches suffer from the issue of excessive low-quality query graphs in expressing a complex question. To address this issue, we propose a comprehensive approach to generating query graphs. We first introduce a model, which leverages contrastive learning as well as features fusion, for core chains reasoning. We next propose a semantic-aware model along with an enhancement strategy to identify and associate constraints on a core chain. Extensive experiments on benchmark datasets show superiority of our approach.