stefan-it's picture
readme: beautify
b3cbe1c
metadata
language: multilingual
license: mit
widget:
  - text: and I cannot conceive the reafon why [MASK] hath
  - text: Täkäläinen sanomalehdistö [MASK] erit - täin
  - text: Det vore [MASK] häller nödvändigt att be
  - text: Comme, à cette époque [MASK] était celle de la
  - text: In [MASK] an atmosphärischen Nahrungsmitteln

hmBERT: Historical Multilingual Language Models for Named Entity Recognition

More information about our hmBERT model can be found in our new paper: "hmBERT: Historical Multilingual Language Models for Named Entity Recognition".

Languages

Our Historic Language Models Zoo contains support for the following languages - incl. their training data source:

Language Training data Size
German Europeana 13-28GB (filtered)
French Europeana 11-31GB (filtered)
English British Library 24GB (year filtered)
Finnish Europeana 1.2GB
Swedish Europeana 1.1GB

Corpora Stats

German Europeana Corpus

We provide some statistics using different thresholds of ocr confidences, in order to shrink down the corpus size and use less-noisier data:

OCR confidence Size
0.60 28GB
0.65 18GB
0.70 13GB

For the final corpus we use a OCR confidence of 0.6 (28GB). The following plot shows a tokens per year distribution:

German Europeana Corpus Stats

French Europeana Corpus

Like German, we use different ocr confidence thresholds:

OCR confidence Size
0.60 31GB
0.65 27GB
0.70 27GB
0.75 23GB
0.80 11GB

For the final corpus we use a OCR confidence of 0.7 (27GB). The following plot shows a tokens per year distribution:

French Europeana Corpus Stats

British Library Corpus

Metadata is taken from here. Stats incl. year filtering:

Years Size
ALL 24GB
>= 1800 && < 1900 24GB

We use the year filtered variant. The following plot shows a tokens per year distribution:

British Library Corpus Stats

Finnish Europeana Corpus

OCR confidence Size
0.60 1.2GB

The following plot shows a tokens per year distribution:

Finnish Europeana Corpus Stats

Swedish Europeana Corpus

OCR confidence Size
0.60 1.1GB

The following plot shows a tokens per year distribution:

Swedish Europeana Corpus Stats

All Corpora

The following plot shows a tokens per year distribution of the complete training corpus:

All Corpora Stats

Multilingual Vocab generation

For the first attempt, we use the first 10GB of each pretraining corpus. We upsample both Finnish and Swedish to ~10GB. The following tables shows the exact size that is used for generating a 32k and 64k subword vocabs:

Language Size
German 10GB
French 10GB
English 10GB
Finnish 9.5GB
Swedish 9.7GB

We then calculate the subword fertility rate and portion of [UNK]s over the following NER corpora:

Language NER corpora
German CLEF-HIPE, NewsEye
French CLEF-HIPE, NewsEye
English CLEF-HIPE
Finnish NewsEye
Swedish NewsEye

Breakdown of subword fertility rate and unknown portion per language for the 32k vocab:

Language Subword fertility Unknown portion
German 1.43 0.0004
French 1.25 0.0001
English 1.25 0.0
Finnish 1.69 0.0007
Swedish 1.43 0.0

Breakdown of subword fertility rate and unknown portion per language for the 64k vocab:

Language Subword fertility Unknown portion
German 1.31 0.0004
French 1.16 0.0001
English 1.17 0.0
Finnish 1.54 0.0007
Swedish 1.32 0.0

Final pretraining corpora

We upsample Swedish and Finnish to ~27GB. The final stats for all pretraining corpora can be seen here:

Language Size
German 28GB
French 27GB
English 24GB
Finnish 27GB
Swedish 27GB

Total size is 130GB.

Pretraining

Details about the pretraining are coming soon.

Acknowledgments

Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC) program, previously known as TensorFlow Research Cloud (TFRC). Many thanks for providing access to the TRC ❤️

Thanks to the generous support from the Hugging Face team, it is possible to download both cased and uncased models from their S3 storage 🤗