OrderSum: Reading Order-Aware Unsupervised Opinion Summarization

Anonymous

Abstract: Opinion summarization aims to create a concise summary reflecting subjective information conveyed by multiple user reviews about the same product. To avoid the high expense of curating golden summaries for training, many unsupervised methods have been recently developed. Most state-of-the-art methods utilize the extracted segments following their salience ranking as pseudo labels to train a summary generator. However, the extracted salient segments can be verbose and their reading order has been long overlooked. In this paper, we propose a reading order-aware framework, OrderSum, aiming to generate concise and logical summaries. Specifically, we first formulate the segment ordering problem in pseudo labels as path-choosing and solve it using reinforcement learning. Moreover, to generate a more concise summary, we propose to encourage the generative model to skip useless words based on the token link information derived from concise sentences, which can be collected easily from massive raw reviews by considering the ratio of sentiment/aspect words. Extensive experiments demonstrate that OrderSum benefits from the awareness of reading order and the conciseness modeling, thus being more effective than existing unsupervised methods and achieving the state-of-the-art performance.