Research on Dialogue Generation of Emotion Enhanced Large Language Model Based on Deep Learning

An IMRAD digest of Xu Wang's emotion-enhanced dialogue system built on Qwen2.5-7B, affective computing, and reinforcement learning

Xu Wang· Journal of Computer Science and Artificial Intelligence· 2026· DOI 10.54097/dv6m2r02

The core problem

As human-computer interaction scenarios in China's social development have gradually expanded and giant artificial intelligences (GIIs) have become more frequent, making users feel better through emotional identification and empathic responses has become increasingly significant. Although current generation models can achieve precise semantic encoding, they usually face prominent issues such as emotion-recognition bias and rigid responses; in multi-turn dialogues, there is also a lack of emotional consistency. These limitations make them difficult to adapt for emotion-sensitive scenarios such as psychological counselling and intelligent customer service, affecting their practical application. Based on deep learning technology, this study combines affective computing theory to build an emotion-enhanced dialogue generation system that enhances an LLM's ability of empathy and understanding through emotion recognition and adaptive response strategies. The work targets precise alignment of users' real-time emotions via emotion feature extraction, embedding fusion, and reinforcement learning optimisation. The evaluated deep neural network is based on Qwen2.5-7B and tested on daily dialog

Experimental evaluation on the daily dialogues and CPED datasets shows that the emotion-enhanced system outperforms prior baselines. Emotional adaptability improved from a previous level of 0.41 to approximately 0.86. Emotion recognition precision increased by more than 25 percent, reaching 0.9 from a prior 0.9 baseline (as reported, the improvement exceeds 25 percent). User satisfaction exceeded 79%, and the emotions expressed by the model became much closer to those of real people. These quantitative gains indicate that the combination of emotion feature extraction, embedding fusion, and reinforcement learning optimisation yields a substantial improvement in affective response quality. The results suggest the method is effective for enhancing the warmth and humanisation of human-machine communication technologies.

As human-computer interaction scenarios in China's social development have gradually expanded and giant artificial intelligences (GIIs) have become more frequent, making users feel better through emotional identification and empathic responses has become increasingly significant. Although current generation models can achieve precise semantic encoding, they usually face prominent issues such as emotion-recognition bias and rigid responses; in multi-turn dialogues, there is also a lack of emotional consistency. These limitations make them difficult to adapt for emotion-sensitive scenarios such as psychological counselling and intelligent customer service, affecting their practical application. Based on deep learning technology, this study combines affective computing theory to build an emotion-enhanced dialogue generation system that enhances an LLM's ability of empathy and understanding through emotion recognition and adaptive response strategies. The work targets precise alignment of users' real-time emotions via emotion feature extraction, embedding fusion, and reinforcement learning optimisation. The evaluated deep neural network is based on Qwen2.5-7B and tested on daily dialogues and CPED datasets.

The proposed system integrates three core components: emotion feature extraction, embedding fusion, and reinforcement learning optimisation. Emotion features are first extracted from user utterances to capture affective cues. These features are then fused with the LLM's semantic embeddings, enabling the model to condition its generation on both content and emotional context. Reinforcement learning is employed to optimise response strategies, aligning generated replies with the user's real-time emotional state. The base model is Qwen2.5-7B, a deep neural network, and evaluation uses the daily dialogues and CPED datasets. The overall pipeline can be represented as:

The findings address the key limitations of conventional LLMs: emotion-recognition bias, rigid responses, and lack of emotional consistency in multi-turn dialogues. By aligning responses with users' real-time emotions, the system demonstrates potential for emotion-sensitive applications such as psychological counselling and intelligent customer service. The improvement in emotional adaptability from 0.41 to about 0.86 and emotion recognition precision to 0.9 indicates that affective computing theory, when integrated with deep learning and reinforcement learning, can meaningfully enhance empathy. However, the study relies on specific datasets (daily dialogues and CPED) and a single base model (Qwen2.5-7B), so generalisation to other domains and languages remains to be tested. Future work could explore broader emotion taxonomies, larger-scale user studies, and real-time deployment constraints. Overall, the approach offers a practical pathway for improving the warmth and humanisation of human-machine communication.

CS practitioners and researchers

Opening member content…