Data Mining and Text Analytics in Music Audi Sugianto and Nicholas Tawonezvi
Overview Introduction Building a ground truth set Experiments Results
Introduction Purpose: Music mood classification through lyric text mining approaches MIR (Music Information Retrieval) Use of Audio Datasets: AMC (Audio Mood Classification) USPOP, USCRAP, etc. Use of Social tags from last.fm Challenges: Natural subjectivity of music Human perspectives on mood
Generating Ground Truth Data Collection Combination of in-house and public audio tracks Collect songs with at least one social tag from last.fm Lyrics can be gathered from mainly Lyricwiki.org. Use of Lingua to ensure data quality Finalise songs that have both correct lyrics and tags
Generating Ground Truth Algorithms, Resources and Techniques WordNet-Affect Used to filter out junk tags Assignment of labels to concepts (emotions, moods, responses) Use of human expertise to identify mood-related words in the music domain Affective Aspect Judgemental Tags Ambiguous Meanings Use of WordNet to categorise into groups based on synonyms. Use of music experts to merge groups by musical similarity
Generating Ground Truth Selecting Songs Approaches: Tag identification Lyric counts Multi-label Classification
Mood Categories and Song Distributions
Experiments Evaluation Measures and Classifiers Use of 10-fold Cross Validation Break data into 10 sets of size n/10. Train on 9 datasets and test on 1. Repeat 10 times and take a mean accuracy. Classification with Support Vector Machines (SVM) Algorithms to analyse data and recognise patterns
Experiments Lyric Preprocessing Facts: Repetitions of words and sections: - Lack of verbatim transcripts Consisting of sections: Intro, interlude, verse, etc. in the annotations Notes about song and instrumentation Possible solution: Identifying and converting repetition and annotation patterns to actual repeated segments
Experiments Lyrics Features Common text classification tasks: Bag-of-words (BOW) Collection of Unordered words Part-of-Speech (POS) Use of Stanford Tagger Function Words (the, a, etc.) Assigning of values: Frequency Tf-idf weight Normalised-frequency Boolean Value
Experiments Stemming Stemming – Merging words with same morphological roots Snowball Stemmer Irregular nouns and verbs as inputs
Results Text categorisation provides dimensionality and good generalisability POS Boolean representation is poorer because of high content of POS types in lyrics Content words are more useful in mood classification 10th International Society for Music Information Retrieval Conference (ISMIR 2009)
Acknowledgement Hu, X. et al Lyric Text Mining in Music Mood Classification. International Music Information Retrieval Systems Evaluation Laboratory University of Illinois at Urbana- Champaign. [Online]. Pp [Accessed 6 December 2013]. Available from ː Training and Testing Data Sets Training and Testing Data Sets. [Online]. [Accessed 5 December 2013]. Available from: Kohavi, Ron (1995). A study of cross-validation and bootstrap for accuracy estimation and model selection. Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence 2 (12): 1137–1143.(Morgan Kaufmann, San Mateo, CA) D. Ellis, A. Berenzweig, and B. Whitman: The USPOP2002 Pop Music Data Set. Available fromː
Software & Additional Resources – Statistical language identifier irregular verb list - irregular noun list POS Tagger - Mood Categories & Song Distributions Tests&pid=1087Tests&pid=1087 – Performance identifier