In e-commerce search, the semantic relevance model assesses the relevance between user search queries and candidate product titles to determine which products can be presented to users. To balance efficiency and accuracy, in industrial scenarios, most current methods adopt relevance models based on late-interaction. However, these methods do not effectively model the fine-grained relevance between query and title, especially for long-tailed words. In this paper, we propose the Interpretable Word Representation Learning Framework, a novel relevance model in e-commerce, which overcomes this challenge and improves the accuracy of relevance calculation. Specifically, to enhance the model’s ability to determine the semantic relevance of word pairs, we propose the sparse alignment based on optimal transport (OT), which focuses on one-way semantic alignment from query to title when calculating relevance. To enhance the model’s ability to represent long-tail words, we combine character-level and word-level semantic representations to encode word information. In addition, we design an effective contrastive learning method to train our model to alleviate the problem of insufficient negative samples. Extensive experiments have verified the excellent performance of our proposed method. We have successfully deployed it to our online search system.