Working Notes for Norwegian BERT-Like Models
Report on the creation of FinBERT: https://arxiv.org/pdf/1912.07076.pdf
Available Bokmål Text Corpora
- Norsk Aviskorpus; 1.7 billion words; sentences are ordered; clean;
- Norwegian Wikipedia; 160 million words; sentences are ordered; clean (more or less);
NAK Bøkmal: 712 145 669 word tokens
NAK Nynorsk: 47 180 985 word tokens
NAK unspecified: 1 119 744 725 word tokens
We start with training a joint BERT model. Meanwhile, we will run langid on unspecified texts to later train separate models.
We currently do not use
- noWAC; 700 million words; sentences are ordered; semi-clean;
- CommonCrawl from CoNLL 2017; 1.3 billion words; sentences are shuffled; not clean;
- NB Digital; 200 million words; sentences are ordered; semi-clean (OCR quality varies).
Special stuff for Norsk Aviskorpus
1. Post-2011 archives contain XML files, one document per file, UTF-8 encoding. A simple XML reader will extract text from them easily. No problems here.
2. For years up to 2005 ('1998-2011/1/' subdirectory), the text is in the one-token-per-line format. There are special delimiters signaling the beginning of a new document and providing the URLs. Will have to decide on how exactly to convert it to running text. Moses de-tokenizer? DONE using a self-made tokenizer (in the repo)
3. Everything up to and including 2011 ('1998-2011/' subdirectory) is in the ISO 8859-01 encoding ('Latin-1'). The '1998-2011/3' subdirectory contains XML files which are in 8859-01 as well, although some of them falsely claim (in their headers) to be UTF-8. Must convert to UTF-8 before any other pre-processing. DONE
Preprocessing and Tokenization
1. Check quotes after detokenization DONE
2. Obtain Nynorsk Wikipedia (since we are going to train a joint BERT model) DONE
3. Find out how Stanza evaluates sentence segmentation. Compare its performance with UDPipe and Punkt. (POSTPONED)
4. Train a joint Stanza sentence segmenter on Nynorsk and Bøkmal if necessary (are embeddings needed?). (POSTPONED)
5. Sentence-segment the corpora. DONE
SentencePiece library finds 157 unique characters in Norwegian Wikipedia dump.
Should we assume that the input to the trained model will be tokenized text (punctuation marks separated from words) or not?
This is an issue of balancing between the needs of a naive user (who wants to avoid any pre-processing) and the needs of a computational linguist (who arguably wants to have more linguistically meaningful tokens at the output).
We decided that the default model is trained on raw text, but if time allows, a `tokenized' model should be trained for comparison.
Training 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). Will do sentence-splitting with Stanza.
2. Blank lines between documents. Document boundaries are needed so that the "next sentence prediction" task doesn't span between documents.
3. Text files are converted to TFRecords. Each TFR is about 60 times larger than the original gzipped text file. We need about 300 GB to store TFRs for sequence length 128 for our full training corpus.
Do we have available Norwegian test sets for typical NLP tasks to evaluate our NorBERT? Please see Eosc/norbert/benchmark for a discussion.