Learning from Adjective-Noun Pairs: A Knowledge-enhanced Framework for Target-Oriented Multimodal Sentiment Classification

Proceedings of the 29th International Conference on Computational Linguistics, pages 6784–6794 October 12–17, 2022. 6784 Learning from Adjective-Noun Pairs: A Knowledge-enhanced Framework for Target-Oriented Multimodal Sentiment Classification Fei Zhao Zhen Wu∗ Siyu Long Xinyu Dai Shujian Huang Jiajun Chen National Key Laboratory for Novel Software Technology, Nanjing University, China Collaborative Innovation Center of Novel Software Technology and Industrialization, China {zhaof, longsy}@smail.nju.edu.cn {wuz, daixinyu, huangsj, chenjj}@nju.edu.cn Abstract Target-oriented multimodal sentiment classi- fication (TMSC) is a new subtask of aspect- based sentiment analysis, which aims to de- termine the sentiment polarity of the opinion target mentioned in a (sentence, image) pair. Recently, dominant works employ the attention mechanism to capture the corresponding visual representations of the opinion target, and then aggregate them as evidence to make sentiment predictions. However, they still suffer from two problems: (1) The granularity of the opinion target in two modalities is inconsistent, which causes visual attention sometimes fail to cap- ture the corresponding visual representations of the target; (2) Even though it is captured, there are still significant differences between the visual representations expressing the same mood, which brings great difficulty to senti- ment prediction. To this end, we propose a novel Knowledge-enhanced Framework (KEF) in this paper, which can successfully exploit adjective-noun pairs extracted from the image to improve the visual attention capability and sentiment prediction capability of the TMSC task. Extensive experimental results show that our framework consistently outperforms state- of-the-art works on two public datasets. 1 Introduction Target-oriented multimodal sentiment classification (TMSC) is a new sub-task of aspect-based senti- ment analysis (Pang et al., 2008; Liu, 2012; Pontiki et al., 2014), which aims to predict the sentiment polarity of the opinion target mentioned in a pair of sentence and image. The assumption behind this task is that the image information can help the text content identify the sentiment of the opinion tar- get. Fig. 1(a) and Fig. 1(b) show two representative examples. It is hard to detect the sentiment of the opinion target (i.e., “Vince Gilligan” or “Sammy”) only depending on informal sentences, but the vi- ∗ Corresponding author. (b) # OOTD with my little dog by my side . [Sammy]positive . issuse1 issuse2 (d) # OOTD with my little dog by my side . [Sammy]positive . positive ? ✘ ✓ (a) [Vince Gilligan]positive travels in the city of cape town . (c) [Vince Gilligan]positive travels in the city of cape town . ✘ ✓ positive ? Figure 1: Two examples of Target-Oriented Multimodal Sentiment Classification (TMSC). Opinion targets and their corresponding sentiment polarities are highlighted in the sentence. The red bounding box denotes the visual clues that the opinion target focuses on. sual content (i.e., smiling face) associated with the target can clearly reflect its sentiment polarity. From the above examples, we can conclude that aligning the opinion target of two modalities and capturing helpful visual sentiment features play a critical role in the TMSC task. Given its impor- tance, dominant works employ the attention mech- anism (Bahdanau et al., 2015) to automatically …

Content truncated to 500 words