Improving Question Answering by Commonsense-Based Pre-Training
Wanjun Zhong, Duyu Tang, Nan Duan, Ming Zhou, Jiahai Wang, Jian Yin
TL;DR
This work addresses the challenge of commonsense reasoning in QA by pre-training functions that model direct and indirect relations between concepts using ConceptNet. The pre-trained commonsense scoring is then integrated with standard document-based QA models to improve multi-choice question answering across ARC, SemEval, and OpenBook QA datasets. The method combines concept-level encodings with graph-structured neighbor information, and uses a ranking-based objective with a simple scoring fusion f(a_i)=\alpha f_{doc}(a_i)+\beta f_{cs}(a_i). Overall, the results show that external commonsense knowledge provides complementary signals that improve reasoning, though the authors acknowledge limitations in disambiguation and complex logical reasoning, proposing context-aware and semantic-parsing enhancements for future work.
Abstract
Although neural network approaches achieve remarkable success on a variety of NLP tasks, many of them struggle to answer questions that require commonsense knowledge. We believe the main reason is the lack of commonsense \mbox{connections} between concepts. To remedy this, we provide a simple and effective method that leverages external commonsense knowledge base such as ConceptNet. We pre-train direct and indirect relational functions between concepts, and show that these pre-trained functions could be easily added to existing neural network models. Results show that incorporating commonsense-based function improves the baseline on three question answering tasks that require commonsense reasoning. Further analysis shows that our system \mbox{discovers} and leverages useful evidence from an external commonsense knowledge base, which is missing in existing neural network models and help derive the correct answer.