Fine-grained Sentiment Controlled Text Generation Approach Based on Pre-trained Language Model

Anonymous

Abstract: Sentiment-controlled text generation aims to generate texts according to the given sentiment. However, most of the existing studies focus only on document- or sentence-level sentiment control, leaving a gap for finer-grained control over the content of generated results. Some previous works attempted to generate reviews conditioned on the aspect-level sentiments, but they usually suffer from low adaptability and the lack of annotated dataset. To alleviate these problems, we propose a pre-trained model extended generative model together with an auxiliary classifier to perform training on both annotated and unannotated datasets. We also propose a query-hint mechanism to further guide the generation process towards the aspect-level sentiments at every time step. Experimental results from real-world datasets demonstrated that our model has excellent adaptability in generating aspect-level sentiment controllable review texts with high sentiment coverage and stable quality.

Paper Type: long