SANCL: Multimodal Review Helpfulness Prediction with Selective Attention and Natural Contrastive Learning

Proceedings of the 29th International Conference on Computational Linguistics, pages 5666–5677 October 12–17, 2022. 5666 SANCL: Multimodal Review Helpfulness Prediction with Selective Attention and Natural Contrastive Learning Wei Han♣∗, Hui Chen♣, Zhen Hai†⋄, Soujanya Poria♣, Lidong Bing† ♣ Singapore University of Technology and Design, Singapore † DAMO Academy, Alibaba Group {wei_han, hui_chen}@mymail.sutd.edu.sg sporia@sutd.edu.sg {zhen.hai, l.bing}@alibaba-inc.com Abstract With the boom of e-commerce, Multimodal Re- view Helpfulness Prediction (MRHP), which aims to sort product reviews according to the predicted helpfulness scores has become a re- search hotspot. Previous work on this task fo- cuses on attention-based modality fusion, in- formation integration, and relation modeling, which primarily exposes the following draw- backs: 1) the model may fail to capture the really essential information due to its indiscrim- inate attention formulation; 2) lack appropriate modeling methods that take full advantage of correlation among provided data. In this pa- per, we propose SANCL: Selective Attention and Natural Contrastive Learning for MRHP. SANCL adopts a probe-based strategy to en- force high attention weights on the regions of greater significance. It also constructs a con- trastive learning framework based on natural matching properties in the dataset. Experimen- tal results on two benchmark datasets with three categories show that SANCL achieves state- of-the-art baseline performance with lower memory consumption. Our implementation code for this paper can be found at https: //github.com/declare-lab/SANCL. 1 Introduction It is unbelievable to witness an e-commerce boom that has transpired over the past decades (Vulkan, 2020; Poria et al., 2020). In the virtual bazaar, countless deals are made between mutu- ally invisible sellers and customers from time to time, under the administrator’s supervision. For customers, it may be their biggest headache to de- termine whether they should pay for a good when being overwhelmed by tempting advertisements, as they know quite a little information about a prod- uct in face of the seller’s meticulous promotions ∗ This work was done when Wei was an intern at Alibaba DAMO Academy. ⋄ Corresponding author. without any external references. In this situation, reviews in e-shops that can provide justification information, are thus of great value to customers. However, the quality of reviews under a certain product page can be disparate—many customers are willing to leave informative feedback on the product, while many others arbitrarily write a few words and even paste irrelevant messages in their comments. Therefore, from the perspective of on- line shopping platforms, they would be welcome and attractive to customers if they provide a ser- vice that can intelligently filter and place the most helpful reviews at the top position. The task in the machine learning field to solve this problem is Review Helpfulness Prediction (RHP) (Tang et al., 2013). With the thriving of multimodal learning (Uppal et al., 2022; Hazarika et al., 2022) research and the handy accessibility of multimodal data in this Internet era, the latest progress incorporated image (vision modality) information into the review help- fulness prediction (RHP) (Liu et al., 2021) as Mul- timodal RHP (MRHP). Although previous work attained …

Content truncated to 500 words