Unsupervised aspect-based summarization using variational autoencoders

Digital platforms have become the primary medium for the public to access content and express feedback. Aspect-level summarization targeting content themes aims to delve deep into the multidimensional structure of the content and understand the intrinsic relationship between the content and public feedback. Existing aspect-level summarization methods predominantly match and summarize based on aspects extracted from the feedback. Such approaches might focus on points that are unrelated to the content itself, failing to genuinely reflect the profound structure of the content theme. To address this, this paper introduces a Content-Topic Aligned Aspect-Level Comment Summarization Model (CTACSum) based on Variational Autoencoders (VAEs). Initially, the model extracts thematic graphs representing different content themes from a content keyword co-occurrence graph constructed based on cognitive map theory. To effectively match content themes and feedback, the model employs VAEs to learn latent variable representations capturing key semantic information from the feedback. It then predicts the relationship between content themes and feedback based on the latent variable representation of the feedback and the features of the content theme graph encoded by Graph Convolutional Networks (GCNs). Ultimately, the model feeds representative feedback of a specific content theme into the VAE to generate a summary encapsulating salient semantic information. To validate the efficacy of the proposed model, we constructed and released a benchmark dataset, RNC, using news-feedback as the corpus. The results demonstrate that the CTACSum model outperforms baseline models in terms of relevance and consistency.