Inferring Chord Sequence Meanings via Lyrics: Process and Evaluation

We improve upon our simple approach for learning the “associational meaning” of chord sequences from lyrics based on contingency statistics induced over a set of lyrics with chord annotations. Specifically, we refine this process by using word alignment tools developed for statistical machine translation, and we also use a much larger set of chord annotations. In addition, objective evaluation measures are included. Thus, this work validates a novel application of lexicon induction techniques over parallel corpora to a domain outside of natural language learning. To confirm the associations commonly attributed to major versus minor chords (i.e., happy and sad, respectively), we compare the inferred word associations against synonyms reflecting this dichotomy. To evaluate meanings associated with chord sequences, we check how often tagged chords occur in songs labeled with the same overall meaning.