Impact of Semantic Similarity on Genre Classification Performance in Movie Descriptions

The objective of this study is to develop a system that automatically categorizes English movie descriptions into six distinct genres: Action, Comedy, Romance, Science Fiction, Drama, and Animation. The dataset, characterized by a balanced distribution across these genres, was processed using natural language processing (NLP) techniques. In the initial phase, pre-trained FastText word embedding vectors were employed to filter out words exhibiting low semantic similarity within the descriptions. This step aims to preserve context-specific content relevant to each genre, thereby enhancing the quality of the textual data. Subsequently, the numerical representation of the texts was obtained using the Term Frequency–Inverse Document Frequency (TF-IDF) method, which is known to improve the performance of classification algorithms by reducing the influence of frequently occurring but less informative words. The resulting vectors were classified using Softmax Regression, Multinomial Naive Bayes, Support Vector Machines, Random Forest, and Multilayer Perceptron models, achieving accuracy rates of 89.07%, 89.75%, 89.75%, 89.75%, 85.40%, and 90.02%, respectively. The integration of FastText-based semantic filtering with TF-IDF vectorization proved to be an effective approach for multi-class text classification tasks. Moreover, an empirical comparison of different machine learning algorithms provides valuable insights for model selection. The findings suggest that semantically simplified descriptions contribute positively to model performance. In this regard, the study makes a significant contribution to text preprocessing in genre classification tasks.