Commit messages concisely describe code changes in natural language and are important for program comprehension and maintenance. Previous studies proposed some approaches for automatic commit message generation, but their performance is limited due to inappropriate representation of code changes and improper combination of translation-based and retrieval-based approaches. To address these problems, this paper introduces a novel framework named COME, in which modification embeddings are used to represent code changes in a fine-grained way, a self-supervised generative task is designed to learn contextualized code change representation, and retrieval-based and translation-based methods are combined through a decision algorithm. The average improvement of COME over the state-of-the-art approaches is 9.2% on automatic evaluation metrics and 8.0% on human evaluation metrics. We also analyse the effectiveness of COME's three main components and each of them results in an improvement of 8.6%, 8.7% and 5.2%.