Abstract:Existing sentiment analysis models mainly rely on evident emotive words within phrases. When the apparent emotional words within phrases are eliminated, the performance of these models will inevitably decrease. The implicit communication of emotion without the use of explicit emotional phrases is highly widespread in several cultures. As a result, a classification model is required to learn the link between contexts and the emotions they trigger in an automatic way. Based on whether the sentence should be segmented at the keyword position, existing methods apply either segmented or nonsegmented approaches. When emotional words are removed from a sentence, the nonsegmented approaches may lose syntactic information. To address these issues, an interactive iapsule network was proposed in this paper to extend the segmented approach. Taking the keyword as the segmented position, the network initializes two BERT models from a pretrained checkpoint with shared parameters as the encoder to process both contexts separately. By using both interactive attention and the capsule network with a dynamic routing algorithm, the model can automatically learn the insightful relationship between the former and the latter contexts. After fusing the former and latter context features, the interactive capsule network leverages both local and global attention to complete the sentiment analysis task. Experimental results on both English and Chinese corpora show that the proposed interactive attention model achieves a better performance than existing methods during implicit sentiment analysis tasks. In addition, the proposed model outperformed the top 3 models on WASSA-2018 implicit English shared tasks.

Interpretable Multimodal Capsule Fusion

Modality-invariant Temporal Representation Learning for Multimodal Sentiment Classification

CMCI: A Robust Multimodal Fusion Method for Spiking Neural Networks

DCMFNet: Deep Cross-Modal Fusion Network for Different Modalities with Iterative Gated Fusion

Bi-Bimodal Modality Fusion for Correlation-Controlled Multimodal Sentiment Analysis

IMF: Interactive Multimodal Fusion Model for Link Prediction

Multimodal Language Analysis with Recurrent Multistage Fusion

Multimodal Sentiment Analysis Based on Cross-Modal Attention and Gated Cyclic Hierarchical Fusion Networks

TransModality: An End2End Fusion Method with Transformer for Multimodal Sentiment Analysis

A transformer-encoder-based multimodal multi-attention fusion network for sentiment analysis

Cross-modality Representation Interactive Learning for Multimodal Sentiment Analysis

An Effective Multimodal Representation and Fusion Method for Multimodal Intent Recognition

Graph Capsule Aggregation for Unaligned Multimodal Sequences

Multimodal Sentiment Analysis Using Multi-tensor Fusion Network with Cross-modal Modeling

Tri-Modalities Fusion for Multimodal Sentiment Analysis

Multimodal Sentiment Analysis Network Based on Distributional Transformation and Gated Cross-Modal Fusion

A Unimodal Reinforced Transformer with Time Squeeze Fusion for Multimodal Sentiment Analysis

A Unimodal Representation Learning and Recurrent Decomposition Fusion Structure for Utterance-Level Multimodal Embedding Learning

Modality to Modality Translation: An Adversarial Representation Learning and Graph Fusion Network for Multimodal Fusion

CapsFusion: Rethinking Image-Text Data at Scale

Interactive capsule network for implicit sentiment analysis