A Controllable Multi-space Feature Disentanglement Framework for Open-Domain Dialogue Generation

Open-domain dialogues are often ambiguous due to the lack of semantic clarity in expression, which is primarily rooted in weak semantic recognition. Current semantic recognition models based on feature disentanglement are designed for single hidden spaces. When the expression contains multiple semantics, disentanglement of the correct features cannot be achieved in such a single hidden space, resulting in inaccurate responses. Besides, the semantics of hidden variables cannot be controlled by sampling in a single hidden space. To address these limitations, we propose a controllable multi-space feature disentanglement framework for open-domain dialogue generation. Our approach disentangles features into different classes of multiple semantic hidden spaces, allowing us to generate dialogue texts with specified semantics by controlling the latent variables. To control the dependence level of the generated text on each category, we further introduce a weighted semantic vector approach to improve text generation. Finally, we conducted comprehensive comparative experiments on the DailyDialog dataset. We analyzed the emotion and action categories of the dataset to validate the performance of our model.