Attention Enhanced BiLSTM for Causal Sentiment Mining in Noisy Social Media Streams

Social media platforms generate a torrent of short, informal messages rich in slang, emojis, and context dependent expressions, making accurate sentiment classification and real time interpretability a major challenge for organizations. We propose a novel CNN BiLSTM MHSA architecture augmented with a Noise Invariant Contrastive Head (NICH) that simultaneously boosts predictive performance and exposes the lexical triggers of sentiment shifts. Our model combines pretrained GloVe embeddings, a lightweight convolutional frontend for local n-gram detection, a bidirectional LSTM for contextual modelling, and a multihead self attention layer to dynamically highlight sentiment bearing tokens. Training proceeds in two phases: embeddings are first frozen for stable convergence, then fine tuned with AdamW under a cosine decay learning rate schedule. On two benchmarks (1.6 M tweets from Sentiment140 and 50 k IMDB reviews), we outperform classical baselines (LR, SVM, RF) nd exceed deep variants lacking attention or bidirectionality. Ablation confirms MHSA as the primary driver of gains followed by bidirectionality and dropout. Attention reveal causal patterns negators, intensifiers, key emojis that underpin predictions, enabling targeted interventions in marketing, crisis response, and customer service. Our framework thus provides a scalable blueprint for causal text mining in any noisy, high velocity data stream where interpretability and timely decisions are paramount