Contradiction detection is a task to recognize contradiction relations between a pair of sentences. Despite the effectiveness of traditional context-based word embedding learning algorithms in many natural language processing tasks, such algorithms are not powerful enough for contradiction detection. Contrasting words such as “overfull” and “empty” are mostly mapped into close vectors in such embedding space. To solve this problem, we develop a tailored neural network to learn contradiction-specific word embedding (CWE). The method can separate antonyms in the opposite ends of a spectrum. CWE is learned from a training corpus which is automatically generated from the paraphrase database, and is naturally applied as features to carry out contradiction detection in SemEval 2014 benchmark dataset. Experimental results show that CWE outperforms traditional context-based word embedding in contradiction detection. The proposed model for contradiction detection performs comparably with the top-performing system in accuracy of three-category classification and enhances the accuracy from 75.97% to 82.08% in the contradiction category.