Classification of Coming‑of‑Age Song Lyrics Using Convolutional Neural Network (CNN) Architecture

Abstract

The coming-of-age theme in song lyrics is rich in emotional expressions, identity transitions, and nostalgia. This study implements a 1-Dimensional (1D) Convolutional Neural Network (CNN) architecture to automatically classify coming-of-age-themed song lyrics. Raw lyrics data were collected through a custom scraping function get_song_lyrics, which then went through a structured text preprocessing stage to remove stopwords and non-semantic components. The cleaned words were represented into a vector space using a pre-trained 50-dimensional GloVe embedding (GloVe 50d) of size (max_length, 50) to capture semantic relationships between words. A 1D CNN model was applied to extract local features in the form of phrase combinations (n-grams) through filter shifting, followed by a Global Max Pooling layer to filter out the most dominant emotional information before the final classification process. As a comparison method and additional analysis, a rule-based approach using TextBlob was applied to extract polarity and subjectivity scores, while the Word Cloud technique was used to visualize the dominance of transitional lexical terms such as grow, leave, and remember. The gap between training and validation performance suggests that the model learned dataset-specific patterns rather than generalized semantic representations. Similar overfitting behavior has been reported in CNN-based lyric and sentiment classification studies when training data are limited or insufficiently diverse. The results showed that the integration of GloVe 50d semantic representation and local feature extraction by 1D CNN was able to produce high and stable accuracy in recognizing the unique characteristics of song lyrics with a maturity theme.