Datasets:

NbAiLab

/

NCC

like 18

Thanks for sharing NCC openly, wonderful initiative!

In the documents from NCC that derive from the newspaper_ocrsource, I notice there are a lot of words that are split in the middle with a whitespace character. They seem to be split at common hyphenation points, e.g. syllables, which suggests that the words probably occurred at the end of a line in a newspaper, and were hyphenated due to layout reasons.

This seems to occur in almost all OCR documents. For example:

Et solid flertall av lærer ne er mot norsk medlemskap i EU. I en meningsmåling som MMI har gjort for Norsk Sko leblad, svarte 52 prosent nei, 22 prosent ja, mens 26 pro sent ikke hadde tatt stand punkt. EU er ventet å bli et hett tema på landsmøtet i lærerla get, som begynner mandag. i det hele tatt bør uttale seg om norsk medlemskap. På spørsmål om Lærerlaget bør ta standpunkt, svarte 75 pro sent nei, 17 prosent ja, mens 8 prosent svarte vet ikke. Dette er i store trekk i samsvar med vedtakene fra årsmøtene i lærerlagets fylkeslag. Av 22 fylkeslag gikk bare fem inn for å fatte et EU-vedtak på landsmøtet. (NTB)

Likely hyphenated words that have a whitespace:

"lærer ne"
"Sko leblad"
"lærerla get"
"pro sent"
"pro sent"
"stand punkt"

At KBLab, we considered running an experiment where a language model was trained on machine translated nb->sv data, and compare it to the performance of models trained on sv data. The languages are close, and translation quality is generally much higher than other language pairs due to the closeness. It was while looking at the output of a translation model that I noticed the whitespaces. As a result lot of the component syllables are being interpreted as individual words.

When performing text searches in your newspaper service ( https://www.nb.no/search?q=%22Et%20solid%20flertall%20av%20l%C3%A6rerne%20er%20mot%22&mediatype=aviser&viewability=ALL ) , most of these hyphenations seem to be joined together correctly. Any insights on why they aren't in NCC?