An Efficient Deep Learning Enabled Multimodal Sentiment Analysis based on Neural Networks and Text Mining Architectures for Short-Form Social Media Data: A Comprehensive Analysis

https://doi.org/10.5281/zenodo.18786020

Abstract

In recent years, massive amounts of short-form video content have been produced on TikTok, Instagram Reels, and YouTube Shorts, among others. This volume of content has changed the digital communications environment once again, this time emphasizing the utter dominance of communications in high-speed, Multimodal, and moving away from the former static and textual presentations. As a result, sentiment analysis, which previously dealt with a single textual and unimodal analysis has, evolved into a system that predicts the intended feeling from the integration of textual, acoustic and visual relationships, termed Multimodal Sentiment Analysis (MSA). Driven by breakthroughs brought about by deep learning applied to MSA, especially with the paradigm shifts introduced by Multimodal Transformers and Cross-Modal mechanisms, among others, the state-of-the-art accuracy record has significantly improved. Advanced as these multimodal learning transformation technologies and architectures are, they incur significant computational costs, especially with the efficiency gap, self-attention, and large parameter counts. The increased costs and computational resources needed have always hindered wide implementation in real-time constrained edge environments. This research addresses two critical challenges simultaneously, and this survey examines these issues with its holistic review covering the period of 2016-Present with the MSA methods. A renewed focus on the articulation of meaningful efficiency in the trade-off in computational costs and system efficiency in the costs of computing systems is a noteworthy focus of this survey. The research proposes a definitive and rigorous set of taxonomies and a range of system architectures, termed efficiency-aware systems. These include, but are not limited to knowledge distillation, model pruning, and low-rank factorization that have been proposed to circumvent the latency challenges of mobile devices. Additionally, we assess both conventional scholarly repositories and newly arrived short-form video collections and highlight the instance distinction between laboratory results and actual field application (in-the-wild). The paper also identifies persistent challenges such as the “missing modality” issue and cross-lingual generalization while elucidating the future research directions with lightweight and robust MSA systems that can harvest high-density social media streams.