flair-lms

This repository is part of my NLP research with flair, a state-of-the-art NLP framework from Zalando Research.

This repository will include various language models (forward and backward) that can be used with flair. It will be updated frequently. So please star or watch this repository 😅

Changelog

September 2019: Multilingual Flair Embeddings trained on JW300 corpus are released.

September 2019: All Flair Embeddings that are now officially available in Flair >= 0.4.3 are listed.

Parameters

All Flair Embeddings are trained with a hidden_size of 2048 and nlayers of 1.

Flair Embeddings

Language model	# Tokens	Forward ppl.	Backward ppl.	Flair Embeddings alias
Arabic	736M	3.39	3.45	ar-forward and ar-backward
Bulgarian (fast)	66M	2.48	2.51	bg-forward-fast and bg-backward-fast
Bulgarian	111M	2.46	2.47	bg-forward and bg-backward
Czech (v0)	778M	3.44	3.48	cs-v0-forward and cs-v0-backward
Czech	442M	2.89	2.90	cs-forward and cs-backward
Danish	325M	2.62	2.68	da-forward and da-backward
Basque (v0)	37M	2.56	2.58	eu-v0-forward and eu-v0-backward
Basque (v1)	37M	2.64	2.31	eu-v1-forward and eu-v1-backward
Basque	57M	2.90	2.83	eu-forward and eu-backward
Persian	146M	3.68	3.66	fa-forward and fa-backward
Finnish	427M	2.63	2.65	fi-forward and fi-backward
Hebrew	502M	3.84	3.87	he-forward and he-backward
Hindi	28M	2.87	2.86	hi-forward and hi-backward
Croatian	625M	3.13	3.20	hr-forward and hr-backward
Indonesian	174M	2.80	2.74	id-forward and id-backward
Italian	1,5B	2.62	2.63	it-forward and it-backward
Dutch (v0)	897M	2.78	2.77	nl-v0-forward and nl-v0-backward
Dutch	1,2B	2.43	2.55	nl-forward and nl-backward
Norwegian	156M	3.01	3.01	no-forward and no-backward
Polish	1,4B	2.95	2.84	pl-opus-forward and pl-opus-backward
Slovenian (v0)	314M	3.28	3.34	sl-v0-forward and sl-v0-backward
Slovenian	419M	2.88	2.91	sl-forward and sl-backward
Swedish (v0)	545M	2.29	2.27	sv-v0-forward and sv-v0-backward
Swedish	671M	6.82 (?)	2.25	sv-forward and sv-backward
Tamil	18M	2.23	4509 (!)	ta-forward and ta-backward

Multilingual Flair Embeddings

I trained multilingual Flair Embeddings on the recently released JW300 corpus. Thanks to half precision support in Flair, both forward and backward Embeddings were trained for 5 epochs for over 10 days. The training corpus has 2,025,826,977 token.

Language model	# Tokens	Forward ppl.	Backward ppl.	Flair Embeddings alias
JW300	2B	3.25	3.37	lm-jw300-forward-v0.1.pt and lm-jw300-backward-v0.1.pt

Notice: this JW300 model is not yet available in a tagged release of Flair, so you need to download it manually:

$ wget https://schweter.eu/cloud/flair-lms/lm-jw300-forward-v0.1.pt
$ wget https://schweter.eu/cloud/flair-lms/lm-jw300-backward-v0.1.pt

Then it can be loaded with:

from flair.embeddings import FlairEmbeddings

jw_forward = FlairEmbeddings("lm-jw300-forward-v0.1.pt")
jw_backward = FlairEmbeddings("lm-jw300-backward-v0.1.pt")

A detailed evaluation on various PoS tagging tasks can be found in this repository.

I would like to thank Željko Agić for providing me access to the corpus (before it was officially released)!

Training tips

This sections covers data collection for training a language model. The next section list some famous corpora and corresponding extraction steps to build a "clean" text corpus.

Corpora

Leipzig Corpora Collection

The Leipzig Corpora Collection provides sentence-segmented corpora for various domains (News, Webcrawl, Wikipedia). Awesome resource!

Europarl: http://www.statmt.org/europarl

The Europarl corpus consists of translated text from the European parliament in several languages. It is very popular in the machine translation community.

Texts for several languages can be found under http://www.statmt.org/europarl. An extracted text can directly be used for training a language model.

Wikipedia

Wikipedia dumps can be found under: https://dumps.wikimedia.org. Just download the <language>wiki-latest-pages-articles.xml.bz2 dump:

wget http://download.wikimedia.org/<language>wiki/latest/<language>wiki-latest-pages-articles.xml.bz2

Sadly it is a xml based dump (and no one really likes xml).

Thus, an extractor tool is needed to extract Wikipedia articles from stupid xml to plain text. I recommend the wiki extractor from https://github.com/attardi/wikiextractor. Just clone that repository or download the tool directly via:

wget https://raw.githubusercontent.com/attardi/wikiextractor/master/WikiExtractor.py

Then run the WikiExtractor using:

python3 WikiExtractor.py -c -b 25M -o extracted <language>wiki-latest-pages-articles.xml.bz2

This will create bz2 archives with plain-text wikipedia articles.

Do extract these smaller chunks just use:

find extracted -name '*bz2' \! -exec bzip2 -k -c -d {} \; > <language>wiki.xml

All wikipedia articles are then stored in <language>wiki.xml. Only four preprocessing steps are needed after extraction:

• Strip some stupid xml tags with: sed -i 's/<[^>]*>//g' <language>wiki.xml
• Remove empty lines with: sed -i '/^\s*$/d' <language>wiki.xml
• Remove the temporarily created extraction folder: rm -rf extracted
• Rename the corpus from xml to txt: mv <language>wiki.xml <language>wiki.txt

Your final cleaned plain-text corpus is then located under <language>wiki.txt.

OpenSubtitles2018

Many subtitles from various movies in several languages could be found on the OPUS webpage.

Just find out the language code for your desired language and download an open subtitles dump:

wget "http://opus.nlpl.eu/download.php?f=OpenSubtitles2018%2F<language>.raw.tar.gz"

Unfortunately, all subtitles are stored in a stupid xml format. But first, extract the archives with:

find . -name '*.gz' -exec gunzip '{}' \;

Then concatenate all xml files into one large xml file:

find OpenSubtitles2018/raw/<language>/ -name *.xml -exec cat {} + > opensubtitles-combined.xml

Instead of using some stupid xml parsing, just throw away all stupid xml tags:

cat opensubtitles-combined.xml | grep -v "<" > opensubtitles-combined.txt

In the final step, just delete the old xml file:

rm opensubtitles-combined.xml

SETimes

The SETimes corpus contains newspaper articles that have been translated into several Balkan languages. The corpus can be found under http://nlp.ffzg.hr/resources/corpora/setimes/. Just extract them and they can be used directly to train a language model.

Just make sure that you have downloaded the plain text corpus and not the Moses tokenized texts.

Tokenization

Notice: do not use tokenized or preprocessed corpora, and do not tokenize or preprocess the data. Plain text is all you need :)

Preparations for training

After you collected all dataset, concatenate it like e.g. cat *.txt >> <language>.dataset.txt.

It is also very important to shuffle your data! This can be done with the shuf command:

cat <language>.dataset.txt | shuf > <language>.dataset.shuffled.txt

Dataset split

One strategy for splitting the shuffled dataset into training, development and test sets could be the following:

• Count the number of lines in your shuffled dataset with: wc -l <language>.dataset.shuffled.txt

• Use 1/500 as development and 1/500 as test set

• Use split command with -l option

• The folder structure must be:

  .
  ├── test.txt
  ├── train
  │   ├── xaa
  │   ├── ...
  │   └── xzz
  └── valid.txt