a game-based approach for collecting semantic music annotations douglas turnbull, rouran liu, luke...
TRANSCRIPT
A Game-Based Approach for Collecting Semantic Music
Annotations
Douglas Turnbull, Rouran Liu, Luke Barrington, Gert Lanckriet
Computer Audition Lab
UC San Diego
ISMIR
September 27, 2007
2
Introduction
Automatic audio content analysis helps to organize, search, recommend, retrieve and describe huge - and growing - music collections
Computer audition systems require significant amounts of high-quality semantic labels for audio content
Collecting this data can be difficult, expensive, slow, boring and inaccurate
If only we could get someone else to do it...
3
Sources of Semantic Information
102
Qua
lity
Quantity101 103 104 105 106
web mine
id3 tags
4
5
Sources of Semantic Information
102
Qua
lity
Quantity101 103 104 105 106
CAL500
human tags
web mine
id3 tags
6
7
Sources of Semantic Information
102
Qua
lity
Quantity101 103 104 105 106
CAL500Pandora
Last.fm
Web mine
id3 tags
Human Computation
8
Human Computation
Many problems that are hard for computers can be easily solved by humans
Many humans spend lots of time solving problems that are of little use
How can we put these “gray cycles” to use?
9
Music Games
Multi-player
Music is social
Music can be subjective
Use group consensus ... but allow personal variations
Collaborative
There are no “right answers”
But agreed-on answers earn more points...
Fun
Need to excite players in order to collect lots of data
Sacrifice data collection in favor of a compelling game
10
www.ListenGame.org
590 players have played at least 1 game
30,000 song-word associations collected
ISMIR deadline
11
Evaluation
Evaluate the quality of collected semantic annotations by using them to train an automatic music retrieval system [SIGIR07]
0.705CAL-250159 Words
0.609AllMusic317 Words
Retrieval ROC AreaDataset0.609AllMusic
317 Words
Retrieval ROC AreaDataset
0.661Listen-25082 Words
0.705CAL-250159 Words
0.609AllMusic317 Words
Retrieval ROC AreaDataset
A Game-Based Approach for Collecting Semantic Music
Annotations
Douglas Turnbull, Rouran Liu, Luke Barrington, Gert Lanckriet
Computer Audition Lab
UC San Diego
ISMIR
September 27, 2007
15
Annotating Music
Web mining
Cheap, collect lots of data
Noisy data, not necessarily related to music content
Surveys / Hand-labelling
e.g. Music Genome Project, LastFM tags, CAL500
Reliable, can be tailored to applications
Expensive, slow, boring, unfocused, free vocabulary
Games
Engage users, free, offer new, social music interaction
Need lots of players!