PGT-FG
A fine-grained multimodal toxicity annotation dataset for Punjabi music content, containing 2,396 segment-level annotations covering binary toxicity, eight fine-grained toxicity categories, severity, sentiment, and modality information.
Timeline (last 90 days)
The researchers constructed the Punjabi Geet Toxicity Dataset (PGT-FG), a fine-grained multimodal toxicity annotation dataset for Punjabi music content.
PGT-FG's data comes from publicly available Punjabi songs on YouTube; the research team initially collected 620 songs and successfully downloaded 465.
PGT-FG ultimately contains 2,396 independent segments from 392 songs.
PGT-FG's annotation was performed by four native Punjabi speakers, producing a total of 2,870 annotation records.
In PGT-FG, 158 segments were covered by all four annotators and were used to compute agreement metrics.
Each annotation in PGT-FG includes a binary toxicity label, eight fine-grained toxicity categories, severity, sentiment orientation, modality information, and an uncertainty flag.
PGT-FG uses Fleiss' kappa to measure inter-annotator agreement on the shared subset, with a kappa value of 0.6521 for the overall toxicity label.
Of the 2,396 segments in PGT-FG, 705 (29.42%) were labeled toxic and 1,691 (70.58%) non-toxic.
Among the toxic segments in PGT-FG, the violence category appears most frequently, involving 539 segments (76.45%), followed by gang references (227, 32.20%) and drug references (133, 18.87%).
PGT-FG has been made public on Zenodo with DOI 10.5281/zenodo.23040156.
2 more timeline events
All Facts (12)
PGT-FG has only four annotators and only 158 segments in the shared annotation subset, so the statistical power of the agreement evaluation is limited.
50%UnverifiedPGT-FG's data comes from publicly available Punjabi songs on YouTube; the research team initially collected 620 songs and successfully downloaded 465.
50%UnverifiedPGT-FG ultimately contains 2,396 independent segments from 392 songs.
50%UnverifiedPGT-FG's annotation was performed by four native Punjabi speakers, producing a total of 2,870 annotation records.
50%UnverifiedIn PGT-FG, 158 segments were covered by all four annotators and were used to compute agreement metrics.
50%UnverifiedEach annotation in PGT-FG includes a binary toxicity label, eight fine-grained toxicity categories, severity, sentiment orientation, modality information, and an uncertainty flag.
50%UnverifiedPGT-FG uses Fleiss' kappa to measure inter-annotator agreement on the shared subset, with a kappa value of 0.6521 for the overall toxicity label.
50%UnverifiedOf the 2,396 segments in PGT-FG, 705 (29.42%) were labeled toxic and 1,691 (70.58%) non-toxic.
50%UnverifiedAmong the toxic segments in PGT-FG, the violence category appears most frequently, involving 539 segments (76.45%), followed by gang references (227, 32.20%) and drug references (133, 18.87%).
50%UnverifiedThe researchers constructed the Punjabi Geet Toxicity Dataset (PGT-FG), a fine-grained multimodal toxicity annotation dataset for Punjabi music content.
50%UnverifiedFor copyright reasons, PGT-FG does not redistribute the original lyrics, lyric excerpts, or copyrighted audio and video content along with the dataset.
50%UnverifiedPGT-FG has been made public on Zenodo with DOI 10.5281/zenodo.23040156.
50%