Aspect-Level Sentiment-Controllable Review Generation with Mutual Learning Framework

Review generation, aiming to automatically generate review text according the given information, is proposed assist in unappealing writing. However, most of existing methods only consider overall sentiments reviews and cannot achieve aspect-level sentiment control. Even though some previous studies attempt sentiment-controllable reviews, they usually require large-scale human annotations which are unavailable real world. To address this issue, we propose a mutual learning framework take advantage unlabeled data generation. The consists generator classifier utilize confidence mechanism reconstruction reward enhance each other. Experimental results show our model can aspect-sentiment control accuracy up 88% without losing generation quality.