This paper introduces a novel model called Tensor Fusion Network (TFN) for multimodal sentiment analysis. The model is designed to capture both intra-modality and inter-modality dynamics in sentiment analysis, particularly for spoken language in online videos along with accompanying gestures and voice. The TFN model outperforms state-of-the-art approaches in both multimodal and unimodal sentiment analysis. Multimodal sentiment analysis involves analyzing text, visual, and acoustic modalities to determine sentiment. The challenge lies in modeling the interactions between these modalities, which can change the perception of sentiment. For example, the utterance "This movie is sick" can be ambiguous, but the sentiment changes based on accompanying gestures or voice. The TFN model consists of three main components: Modality Embedding Subnetworks, Tensor Fusion Layer, and Sentiment Inference Subnetwork. The Modality Embedding Subnetworks process language, visual, and acoustic features to generate rich embeddings. The Tensor Fusion Layer explicitly models unimodal, bimodal, and trimodal interactions using a 3-fold Cartesian product. The Sentiment Inference Subnetwork uses the output of the Tensor Fusion Layer to perform sentiment analysis. The model was evaluated on the CMU-MOSI dataset, which contains video opinions from YouTube movie reviews. The results showed that TFN outperformed previous state-of-the-art approaches in both multimodal and unimodal sentiment analysis. The model also demonstrated that each of the three Modality Embedding Subnetworks outperformed unimodal state-of-the-art approaches. The TFN model is effective in capturing the complex interactions between modalities, leading to improved sentiment analysis performance. The model's ability to handle the volatile nature of spoken language, where proper language structure is often ignored, makes it particularly suitable for multimodal sentiment analysis. The model's performance was validated through extensive experiments, showing its effectiveness in capturing both intra-modality and inter-modality dynamics.This paper introduces a novel model called Tensor Fusion Network (TFN) for multimodal sentiment analysis. The model is designed to capture both intra-modality and inter-modality dynamics in sentiment analysis, particularly for spoken language in online videos along with accompanying gestures and voice. The TFN model outperforms state-of-the-art approaches in both multimodal and unimodal sentiment analysis. Multimodal sentiment analysis involves analyzing text, visual, and acoustic modalities to determine sentiment. The challenge lies in modeling the interactions between these modalities, which can change the perception of sentiment. For example, the utterance "This movie is sick" can be ambiguous, but the sentiment changes based on accompanying gestures or voice. The TFN model consists of three main components: Modality Embedding Subnetworks, Tensor Fusion Layer, and Sentiment Inference Subnetwork. The Modality Embedding Subnetworks process language, visual, and acoustic features to generate rich embeddings. The Tensor Fusion Layer explicitly models unimodal, bimodal, and trimodal interactions using a 3-fold Cartesian product. The Sentiment Inference Subnetwork uses the output of the Tensor Fusion Layer to perform sentiment analysis. The model was evaluated on the CMU-MOSI dataset, which contains video opinions from YouTube movie reviews. The results showed that TFN outperformed previous state-of-the-art approaches in both multimodal and unimodal sentiment analysis. The model also demonstrated that each of the three Modality Embedding Subnetworks outperformed unimodal state-of-the-art approaches. The TFN model is effective in capturing the complex interactions between modalities, leading to improved sentiment analysis performance. The model's ability to handle the volatile nature of spoken language, where proper language structure is often ignored, makes it particularly suitable for multimodal sentiment analysis. The model's performance was validated through extensive experiments, showing its effectiveness in capturing both intra-modality and inter-modality dynamics.