Hierarchical-Gate Multimodal Network for Human Communication Comprehension

Computational modeling of human multimodal language is an emerging research area in natural language processing spanning the language, visual and acoustic modalities. Comprehending multimodal language requires modeling not only the interactions within each modality (intra-modal interactions) but more importantly the interactions between modalities (cross-modal interactions). In this paper, we present a novel neural architecture for understanding human communication called the Hierarchical-gate Multimodal Network (HGMN). Specifically, each modality is first encoded by Bi-LSTM which aims to capture the intra-modal interactions within single modality. Subsequently, we merge the independent information of multi-modality using two gated layers. The first gate which is named as modality-gate will calculate the weight of each modality. And the other gate called temporal-gate will control each time-step contribution for final prediction. Finally, the max-pooling strategy is used to reduce the dimension of the multimodal representation, which will be fed to the prediction layer. We perform extensive comparisons on five publicly available datasets for multimodal sentiment analysis, emotion recognition and speaker trait recognition. HGMN shows state-of-the-art performance on all the datasets.