Exploring Visual Information Enhancement for Multimodal Customized Opinion Generation

We present a novel task called customized opinion generation, which aims to generate a customized review that a specific user would give to a product that has not been yet reviewed by the user. This task can assist users in writing a high-quality review of an unreviewed product. Furthermore, the customized review of a product is useful to reduce the time it takes for users to learn about the product, which ideally should be tailored to the reader. To this end, we study the customized opinion generation task under the approach of pre-trained language models. In particular, we first introduce a visual information enhancement module. In this module, we generate pseudo input words and image captions based on the images, and we introduce various prompt strategies to fully exploit the semantics of the model input text. Besides, we design an interaction model to explicitly capture the influence of input text and product images during customized opinion generation. Experimental results indicate that the proposed model can generate customized reviews, which in many cases are high quality. The results also show that the proposed model gives better results compared to several strong baselines. To our knowledge, we are the first to generate a customized review for an unreviewed product.