Language-Based Syllogistic Reasoning Using Deep Neural Networks

Syllogism is a common form of deductive reasoning that requires precisely two premises and one conclusion. It is considered as a logical method to arrive at new information. However, there has been limited research on language-based syllogistic reasoning that is not typically used in logic textbooks. In support of this new field of study, the authors created a dataset comprised of common-sense English pair sentences and named it Avicenna. The results of the binary classification task indicate that humans recognize the syllogism with 98.16% and the Avicenna-trained model with 89.19% accuracy. The present study demonstrates that aided with special datasets, deep neural networks can understand human inference to an acceptable degree. Further, these networks can be used in designing comprehensive systems for automatic decision-making based on textual resources with near human-level accuracy.