Eosc/norbert

From Nordic Language Processing Laboratory
Revision as of 13:46, 6 November 2020 by Andreku (talk | contribs)
Jump to: navigation, search

Working Notes for Norwegian BERT-Like Models

Report on the creation of FinBERT: https://arxiv.org/pdf/1912.07076.pdf

Working NVIDIA implementation workflow on Saga: /cluster/projects/nn9447k/andreku/NorBERT

Available Bokmål Text Corpora

Preprocessing and Tokenization

SentencePiece library finds 157 unique characters in Norwegian Wikipedia dump.

Input file format:

1. One sentence per line. These should ideally be actual sentences, not entire paragraphs or arbitrary spans of text. (Because BERT uses the sentence boundaries for the "next sentence prediction" task).

2. Blank lines between documents. Document boundaries are needed so that the "next sentence prediction" task doesn't span between documents.

It seems there are some tokenization issues in the UDPipe Norwegian model trained on UD 2.5 (norwegian-bokmaal-ud-2.5-191206.udpipe). It always merges punctuation marks with the preceding words, as can be checked at the online demo.

Example:

# text = På østsiden av vannet, er det en godkjent bålplass.
1	På	på	ADP	_	_	_	_	_	_
2	østsiden	østsid	NOUN	_	Definite=Def|Gender=Masc|Number=Sing	_	_	_	_
3	av	av	ADP	_	_	_	_	_	_
4	vannet,	$vannet,	PUNCT	_	_	_	_	_	_
5	er	være	AUX	_	Mood=Ind|Tense=Pres|VerbForm=Fin	_	_	_	_
6	det	det	PRON	_	Gender=Neut|Number=Sing|Person=3|PronType=Prs	_	_	_	_
7	en	en	DET	_	Gender=Masc|Number=Sing|PronType=Art	_	_	_	_
8	godkjent	godkjent	ADJ	_	Definite=Ind|Degree=Pos|Gender=Neut|Number=Sing	_	_	_	_
9	bålplass.	bålplass.	NOUN	_	Abbr=Yes	_	_	_	SpaceAfter=No

The model trained on the previous 2.4 release (norwegian-bokmaal-ud-2.4-190531.udpipe) does not exhibit such behavior.

There is an active pull request which supposedly fixes this.

Evaluation

Do we have available Norwegian test sets for typical NLP tasks to evaluate our NorBERT?