A sample of descriptions in the dataset: https://gist.github.com/Vesihiisi/1925cecc8b1d94c8a7bcd1d90fe74b2f

A suggestion of how they could be handled:

In order to be included, a description must:

• Not start with ämne; if it does, it's excluded.
• Not end with a full stop; final full stop is removed
• Be at most x (8?) words long – to exclude essays like Farmaceut, teaterman, fotläkare; dansk cand. pharm., frivillig i slesvig-holsteinska kriget 1848-1850, chef för Stockholms stadsbudskår (från 1863), teaterdirektör, liktornsoperatör. Död i Bergen. Longer descriptions are excluded as it's difficult to shorten them elegantly. Splitting on a full stop and treating the first one as a sentence doesn't work with abbreviations like med. lic.
• Start with a lowercase letter, as per Wikidata standards. Thus conversion is done.

(Ok, deep inside I just kinda want to ditch everything longer than like 3 words and containing any punctuation, but let's try :/)

Result:

• Brittisk historiker. → brittisk historiker
• Verksam vid institutionen för ekologi, miljö och geovetenskap, Umeå universitet (2017) → excluded due to >8 words
• Ämnen: miljö och hållbar utveckling. Givit ut diktsamling. → excluded due to blacklisted word

This logic could benefit from an extra pair of eyes, so ping @Lokal_Profil

Now that I have noticed that ämne(n): can also occur in the middle of the description, it should be searched for anywhere in the string (and disqualify it).

Taken a quick look/stab at this. I think your gut instinct is right and 3-5 words is a more reasonable than 8 words.

I think the descriptions can be used if:

• any string containing one of the following is discarded 'ämne', 'birthday.se', 'lc auth'* (lower case comparison).
• any string with more than 4** words is discarded
• any string not starting with an alphabet character is discarded (i.e. numbers, punctuation etc.)
• any initial NE: or DB: strings are removed
• any trailing full stop is removed
• the first character is converted to lower case***

I put a gist of the cleaned results and the cleaning script at https://gist.github.com/lokal-profil/37547b9925a5489dbc6d533fd0be662c . Out of the 1933 strings 977 could be salvaged which isn't bad.

*There might be more strings which a skim read might reveal.
**When I got down to that length (4 worst) most strings seemed ok. Going to 5 words (also in gist) salvages another 130ish strings. A quick look suggests these might also be ok.
***Converting to the first character lower case gives a few bad results (e.g. TV->tV) but they are rare enough that it should be acceptable.