Type-agnostic and form-oriented deductive conclusion generation.

Deductive Conclusion Generation (DCG) aims to generate logically valid conclusions given a set of premises. Existing DCG models adopt type-dependent approaches, making them exhibit error propagation problems or even are unusable in the absence of reasoning type labels. Additionally, their fact-oriented training methods fail to learn correct reasoning patterns due to neglecting the importance of reasoning forms. In this article, we propose a Type-agnostic and Form-oriented (TaFo) DCG model. TaFo employs a type-agnostic approach to integrate the knowledge of various reasoning types, enabling various types of reasoning even in the absence of type labels. In addition, TaFo learns valid reasoning forms before factual deduction, which improves its capacity for handling both factual and counterfactual deductions. Experimental results on the EntailmentBank and QASC datasets show that TaFo outperforms existing methods. Furthermore, TaFo achieves 25 % performance improvement compared to existing methods when reasoning with counterfactual data.