A Hybrid Bidirectional LSTM Framework for Multilingual Sentiment Analysis of Code-Mixed E-Commerce Reviews

The rapid expansion of digital platforms has led to an unprecedented increase in user-generated textual content, particularly in the form of customer reviews, which serve as a primary medium for expressing consumer opinions. A notable characteristic of this modern data is its inherent multilingualism and the phenomenon of code-switching, which creates critical bottlenecks for traditional monolingual sentiment analysis tools. This research proposes an automated, high-precision framework for Multilingual Sentiment Analysis utilising a Hybrid Bidirectional Long Short-Term Memory (BILSTM) architecture. The study employs a large-scale dataset consisting of 50,000 multilingual reviews from e-commerce platforms. To ensure data integrity, a rigorous 12-step computational pipeline was established, involving the removal of 418 duplicate entries and the application of a custom text transformation function. Advanced preprocessing techniques, including Regex-based tokenisation, Porter Stemming, and script normalisation, were implemented to reduce linguistic noise and consolidate the vocabulary into a 10,000-word index. The core methodology involves mapping these cleaned tokens into a 128-dimensional dense vector space to achieve language-agnostic semantic alignment. The architectural framework utilises stacked BILSTM layers with 128 and 64 units, respectively, optimised via the Adam algorithm and protected against overfitting through Spatial Dropout (0.3) and Early Stopping. Experimental results demonstrate that the proposed BILSTM model achieved a superior test accuracy of 87.82%, significantly outperforming the standard unidirectional LSTM baseline of 85.57%. The model correctly identified 4,334 negative and 4,152 positive reviews from a test subset of 9,917 samples, proving its high discriminative power. Furthermore, the bidirectional gates proved highly effective at capturing sentiment intent in extended reviews reaching lengths of up to 2,525 words. This study concludes that the integrated deep learning pipeline effectively transforms raw, chaotic user feedback into structured, actionable insights, providing a scalable solution for real-time global consumer expression analysis in diverse linguistic environments.