Decoupled variational autoencoder with interactive attention for affective text generation

Artificial emotional intelligence (AEI) is an important ability for intelligent systems. Representing emotional states as a continuous sentiment intensity could achieve a more fine-grained sentiment application than traditional categorical approaches. By using sentiment intensity, the main challenge of the existing variational autoencoder methods lies in distinguishing emotional information from semantic information. Practically, it is difficult to ensure that the input training texts do not contain any sentiment. If the disentangled latent variable for text generation contains sentiment features that conflict with the assigned sentiment, the generated texts will be messy. Therefore, we propose a decoupled variational autoencoder (VAE) with interactive attention to solve this problem. The proposed method applies a sentiment decoupler to extract the sentiment from the text. Then, sentiment embeddings are applied to map the intensities into the latent space. To enhance the representation ability of sentiment embeddings and the performance of the decoder, we consider affective text generation as a process of denoising, design a noisy sampling strategy in training, and then continuously update the emotional information with variational attention through a dynamic update mechanism. Extensive experiments are conducted on the Yelp dataset and the Amazon dataset, and the experimental results show that our method outperforms other VAE-based methods in fine-grained affective text generation.

Automated summary

Artificial emotional intelligence (AEI) is crucial for intelligent systems, and representing emotional states as continuous sentiment intensity enables more precise sentiment application. However, existing variational autoencoder methods face challenges in distinguishing emotional information from semantic information. To address this, we propose a decoupled VAE with interactive attention that extracts and maps sentiment intensities into the latent space using sentiment decouplers and embeddings.