Depictor: Topic-Guided Opinion Summarization for Product Reviews With Dual-Perspective Topic Modeling

Abstract

Opinion summarization aims to refine opinions from large-scale reviews, often using select-then-summary methods. Due to the length limitation of the input, only a small number of samples are usually selected for the summarization model, with the risk of ignoring global opinion information such as product aspects and user sentiments. Topic modeling can unsupervisedly extract topic words from texts, holding the potential for capturing global opinion. Therefore, we propose Depictor, a topic-guided two-stage opinion summarization approach with dual-perspective topic modeling (BERTopic and Latent Dirichlet allocation). The dual-perspective topic modeling extracts topic words from both semantic and statistical perspectives. Then the extracted topic information is incorporated into the generator in two ways. For the input side, the topic words are concatenated with the subset reviews from the extractor as supplementary keyword information. For the representation side, an additional topic-driven attention mechanism focusing on the topic words is added to enable the summarization model to pay extra attention to aspect-related keywords during the generation. Experimental results on AmaSum show that the proposed topic-augmented method outperforms several strong baselines, indicating its effectiveness in opinion summarization.