Cross-domain aspect/sentiment-aware abstractive review summarization by combining topic modeling and deep reinforcement learning

摘要

Review text has been widely studied in traditional tasks such as sentiment analysis and aspect extraction. However, to date, no work is toward the end-to-end abstractive review summarization that is essential for business organizations and individual consumers to make informed decisions. This study takes the lead to study the aspect/sentiment-aware abstractive review summarization in domain adaptation scenario. Our novel model Abstractive review Summarization with Topic modeling and Reinforcement deep learning (ASTR) leverages the benefits of the supervised deep neural networks, reinforcement learning, and unsupervised probabilistic generative model to strengthen the aspect/sentiment-aware review representation learning. ASTR is a multi-task learning system, which simultaneously optimizes two coupled objectives: domain classification (auxiliary task) and abstractive review summarization (primary task), in which a document modeling module is shared across tasks. The main purpose of our multi-task model is to strengthen the representation learning of documents and safeguard the performance of cross-domain abstractive review summarization. Specifically, ASTR consists of two key components: (1) a domain classifier, working on datasets of both source and target domains to recognize the domain information of texts and transfer knowledge from the source domain to the target domain. In particular, we propose a weakly supervised LDA model to learn the domain-specific aspect and sentiment lexicon representations, which are then fed into the neural hidden states of given reviews to form aspect/sentiment-aware review representations; (2) an abstractive review summarizer, sharing the document modeling module with the domain classifier. The learned aspect/lexicon-aware review representations are fed into a pointer-generator network to generate aspect/sentiment-aware abstractive summaries of given reviews by employing a reinforcement learning algorithm. We conduct extensive experiments on real-life Amazon reviews to evaluate the effectiveness of our model. Quantitatively, ASTR achieves better performance than the state-of-the-art summarization methods in terms of ROUGE score and human evaluation in both out-of-domain and in-domain setups. Qualitatively, our model can generate better sentiment-aware summarization for reviews with different categories and aspects.