Exchange about topic modeling etc. at TIB

Attendees

Mila will be product manager of automatic classification at TIB
Asim gives a presentation about TrenDTF from a developer's point of view
- Covering FTX metadata and transformation, full text document extraction (partly OCR)
- Results displayed in Umbiko and (planned) in a DSpace repository
Jacopo
- So far, auto classification done by a black box (Averbis)
- Goal to come up with an alternative for TIB, of course at least of the same quality
Christine
- For now, broad classification (like Averbis) is the minimal goal
Mila
- For Annif, TIB portal is just one use case, if also a very pressing one
- Needs to be replaced, best in June 2022 (or by end of the year)
- Subject librarians need a feedback loop (this is a requirement, cannot be done with Averbis)
- Need to assign conference subject topics
- Annif needs to be checked against GND
- TFI combined with Omikuji was best approach within Annif so far, but might be replaced
- Annif community: Hints that fulltext based approach might by worse than condensed abstract or so
- Pretests for TIB Portal: https://doi.org/10.5281/zenodo.4316549
- Graph which percentages are assigned by which LinSearch stage in the TIB portal: LinSearch