Analysis of IMDb Movie Reviews and Ratings Using a Language Model Framework

Over the past few decades, IMDb (Internet Movie Database) has become prominent for rating and user-generated reviews of movies internationally. However, there has not been much research that evaluates the quality of user-generated reviews and their associated ratings. In this study, we present a comprehensive deep learning-based framework that integrates movie rating prediction, abuse detection, sentiment analysis, and aspect-based review summarisation (defined here as categorizing reviews into predefined aspects such as Storyline, Acting, Cinematography, Soundtrack, and Rewatchability, rather than generating shortened textual summaries) to analyse IMDb movie reviews across different rating tiers. We extracted and curated 137,886 reviews from 8,582 movies and employed pre-trained GloVe embeddings to obtain word vectors for high-frequency vocabulary, facilitating an in-depth examination of how linguistic features vary across ratings. Our findings indicate that higher-rated reviews predominantly express positive sentiment, whereas lower-rated reviews exhibit a higher prevalence of abusive language. We investigate sentiment tendencies across various nationalities and genres using a language model framework, offering insights into cultural narratives and public perception of different film themes. Our results show that even high-rated reviews can exhibit strong abusive language, while low-rated reviews can contain positive sentiments. Across all rating levels, re-watchability consistently emerges as a central theme, highlighting the nuanced and complex nature of user-generated movie reviews.