Enhancing Explainability in Sentiment Analysis Using ConvLSTM with SHAP and LIME: A Case Study on the IMDb Movie Review Dataset

Abstract

This paper presents a sentiment analysis framework that combines Convolutional Long Short-Term Memory (ConvLSTM) networks with Explainable Artificial Intelligence (XAI) techniques to improve transparency and interpretability in text classification tasks. Using the IMDb movie review dataset, the study replicates previous ConvLSTM-based sentiment analysis models and extends them by incorporating explainability methods such as SHAP and LIME to provide insight into model decision-making processes. The proposed model achieved an accuracy of 88.5% on the test dataset, demonstrating strong capability in distinguishing between positive and negative sentiments in textual reviews. Furthermore, SHAP and LIME analyses revealed that words such as "poor," "bland," "bad," and "boring" contributed significantly to negative predictions, while words including "interesting," "enjoyed," and "moving" had strong positive influence on classification outcomes. The integration of XAI techniques addressed the interpretability limitations commonly associated with deep learning models, making the ConvLSTM framework more transparent, understandable, and trustworthy for end users. In addition, the study highlights the importance of explainability in educational and research applications where model transparency is essential. This research contributes to the growing field of explainable sentiment analysis by validating interpretability approaches that align closely with human reasoning and enhance confidence in machine learning predictions.