Unsupervised Abstractive Opinion Summarization by Generating Sentences with Tree-Structured Topic Guidance

Abstract This paper presents a novel unsupervised abstractive summarization method for opinionated texts. While the basic variational autoencoder-based models assume unimodal Gaussian prior latent code of sentences, we alternate it with recursive mixture, where each mixture component corresponds to topic sentence and is mixed by tree-structured distribution. By decoding component, generate sentences guidance, root conveys generic content, leaf describe specific topics. Experimental results demonstrate that generated are appropriate as summary texts, which more informative cover input contents than those recent model (Bražinskas et al., 2020). Furthermore, variance Gaussians represents granularity analogous word embedding (Vilnis McCallum, 2015).