This paper proposes a dynamical Emotion-Semantic Correlation Model (ESCM) for empathetic dialogue generation. ESCM dynamically constructs emotion-semantic vectors through the interaction of context and emotions. By encoding emotion-semantic vectors, the model dynamically adjusts emotions and semantics in the context to capture the variability of emotional words. To reflect the correlations between emotions and semantics clearly, we introduce a dependency tree. Based on the dynamic emotion-semantic representation and the dependency tree, ESCM proposes a dynamic correlation graph convolutional network to guide the model to capture the correlations between emotions and semantics clearly. By learning dynamic emotion-semantic representations and their correlations, ESCM accurately understands the emotions of the dialogue and captures important semantics to generate more empathetic responses. The ESCM model is evaluated on the EMPATHETIC-DIALOGUES dataset. The results show that ESCM accurately understands the dialogue and generates grammatically fluent and informative empathetic responses. Furthermore, we extract and statistically analyze the common correlation structures in dialogues from the Empathetic-Dialogue dataset. The results indicate that the correlations between emotion and semantics are frequently and extensively utilized in expressing emotions during conversations. Additionally, the results of our analysis of correlation structures are consistent with linguistic conclusions. Our contributions are as follows: (1) We introduce the expressive characteristics of emotions in linguistics, including the variability of emotions and the correlations between emotions and semantics, to enhance the understanding of the meaning in conversations. (2) We propose the ESCM model, which constructs dynamic emotion-semantic vectors to adjust the dynamics of emotions, and leverages a dependency tree-based dynamic correlation graph convolutional network to learn correlations, in order to generate empathetic responses. (3) Experiments on the EMPATHETIC-DIALOGUE dataset demonstrate the effectiveness of ESCM. Furthermore, additional statistical and analytical experiments show that the correlations in dialogue are consistent with psychological research.This paper proposes a dynamical Emotion-Semantic Correlation Model (ESCM) for empathetic dialogue generation. ESCM dynamically constructs emotion-semantic vectors through the interaction of context and emotions. By encoding emotion-semantic vectors, the model dynamically adjusts emotions and semantics in the context to capture the variability of emotional words. To reflect the correlations between emotions and semantics clearly, we introduce a dependency tree. Based on the dynamic emotion-semantic representation and the dependency tree, ESCM proposes a dynamic correlation graph convolutional network to guide the model to capture the correlations between emotions and semantics clearly. By learning dynamic emotion-semantic representations and their correlations, ESCM accurately understands the emotions of the dialogue and captures important semantics to generate more empathetic responses. The ESCM model is evaluated on the EMPATHETIC-DIALOGUES dataset. The results show that ESCM accurately understands the dialogue and generates grammatically fluent and informative empathetic responses. Furthermore, we extract and statistically analyze the common correlation structures in dialogues from the Empathetic-Dialogue dataset. The results indicate that the correlations between emotion and semantics are frequently and extensively utilized in expressing emotions during conversations. Additionally, the results of our analysis of correlation structures are consistent with linguistic conclusions. Our contributions are as follows: (1) We introduce the expressive characteristics of emotions in linguistics, including the variability of emotions and the correlations between emotions and semantics, to enhance the understanding of the meaning in conversations. (2) We propose the ESCM model, which constructs dynamic emotion-semantic vectors to adjust the dynamics of emotions, and leverages a dependency tree-based dynamic correlation graph convolutional network to learn correlations, in order to generate empathetic responses. (3) Experiments on the EMPATHETIC-DIALOGUE dataset demonstrate the effectiveness of ESCM. Furthermore, additional statistical and analytical experiments show that the correlations in dialogue are consistent with psychological research.