notesum.ai

Published at October 21

PODTILE: Facilitating Podcast Episode Browsing with Auto-generated Chapters

q-bio.GN

cs.AI

cs.LG

Released Date: October 21, 2024

Authors: Azin Ghazimatin¹, Ekaterina Garmash², Gustavo Penha³, Kristen Sheets⁴, Martin Achenbach¹, Oguz Semerci⁵, Remi Galvez⁶, Marcus Tannenberg⁷, Sahitya Mantravadi⁶, Divya Narayanan⁶, Ofeliya Kalaydzhyan⁵, Douglas Cole⁵, Ben Carterette⁶, Ann Clifton⁶, Paul N. Bennett⁵, Claudia Hauff⁸, Mounia Lalmas²

Aff.: ¹Spotify, Berlin, Germany; ²Spotify, London, UK; ³Spotify, Amsterdam, Netherlands; ⁴Spotify, San Francisco, US; ⁵Spotify, Boston, US; ⁶Spotify, New York, US; ⁷Spotify, Gothenburg, Sweden; ⁸Spotify, Delft, Netherlands

Arxiv: https://arxiv.org/abs/2410.16148v1

Podcast dataset ( $k=45$ )
	Model	Chunk size (words)	Static/Dynamic context	WinDiff $\downarrow$	ROUGEL_F1 $\uparrow$	SBERT_F1 $\uparrow$
(1)	CATS (Somasundaran et al., 2020)	-	-	0.505	-	-
(2)	GPT-4 (Achiam et al., 2023)	-	✓/✗	0.448	0.134	0.315
(3)	Gen (seg+label) (Inan et al., 2022)	8000	✗/✗	0.364	0.208	0.394
(4)	PODTILE	7000+1000	✓/✓	0.365	0.231^‡	0.414^‡
(5)	PODTILE (Ablation)	7000+1000	✓/✗	0.368	0.235^‡	0.418^‡
(6)		7000+1000	✗/✓	0.368	0.209	0.392
(7)		7000	✗/✗	0.371	0.215^‡	0.400^‡
Wikisection dataset ( $k=6$ )
(8)	CATS (Somasundaran et al., 2020)	-	-	0.113	-	-
(9)	Gen (seg+label) (Inan et al., 2022)	8000	✗/✗	0.188	0.873	0.925
(10)	PODTILE	7000+1000	✓/✓	0.134	0.866	0.924
QMSum dataset ( $k=85$ )
(11)	CATS (Somasundaran et al., 2020)	-	-	0.469	-	-
(12)	Gen (seg+label) (Inan et al., 2022)	8000	✗/✗	0.443	0.196	0.326
(13)	PODTILE	7000+1000	✓/✓	0.443	0.234	0.365