Robust Social Media Sentiment Analysis using Feature Tokenizer Transformer Architecture

Abstract

Generational in nature, social is one of the most popular inputs used to obtain a sentiment analysis regarding public opinion, customer satisfaction and event awareness. We were trained on Social Media Data: Given the data heterogeneity, Systematic noise — informal nature with semantically ambiguous features and complex feature interactions have limited classical machine learning approaches. This work therefore proposes a new Feature Tokenizer Transformer (FT-Transformer) framework to overcome these challenges for social media sentiment classification and brand reputation analysis. In our model, we take the structured input attributes and embed them into feature tokens that – by design – can learn local correlation between features and global correlations via multi-head self-attention without manual feature engineering. It is the first work that integrates feature tokenization and transformer-based representation learning using end-to-end evaluation pipelines with complete sets of classical classification, regression, agreement, error and computational performance metrics. Experimental results demonstrate that the proposed FT-Transformer provides an effective and scalable solution for intelligent social media sentiment classification with results consistently exhibiting high reliability, good generalization capability, strong robustness of classification performance, low prediction error and computation behavior.