This work introduces a data-driven approach for assigning emotions to music tracks. Consisting of two distinct phases, our framework enables the creation of synthetic emotion-labeled datasets that can serve both Music Emotion Recognition and Auto-Tagging tasks. The first phase presents a versatile method for collecting listener-generated verbal data, such as tags and playlist names, from multiple online sources on a large scale. We compiled a dataset of 5,892 tracks, each associated with textual data from four distinct sources. The second phase leverages Natural Language Processing for representing music-evoked emotions, relying solely on the data acquired during the first phase. By semantically matching user-generated text to a well-known corpus of emotion-labelled English words, we are ultimately able to represent each music track as an 8-dimensional vector that captures the emotions perceived by listeners. Our method departs from conventional labeling techniques: instead of defining emotions as generic ''mood tags'' found on social platforms, we leverage a refined psychological model drawn from Plutchik's theory, which appears more intuitive than the extensively used Valence-Arousal model.