ylilauta-corpus

Tools and data related to the Ylilauta corpus

Data source: http://urn.fi/urn:nbn:fi:lb-2016101210

Data licence: CC-BY-NC 4.0 (See data/README.txt)

Processing

Download and unpack source data

wget https://korp.csc.fi/download/Ylilauta/ylilauta_20150304.zip
unzip ylilauta_20150304.zip

Reformat to document-per-line format with fastText labels

python3 scripts/reformat.py --dedup --fix-text --strip-refs --min-toks 10 \
        ylilauta_20150304.vrt > ylilauta_20150304.txt

Split into train/dev/test

Get label statistics

cut -d ' ' -f 1 ylilauta_20150304.txt | perl -pe 's/__label__//' \
        | sort | uniq -c | sort -rn > label_stats.txt

Split by label

mkdir split
awk '{ print $2 }' label_stats.txt | while read l; do
    egrep "^__label__$l " ylilauta_20150304.txt > split/$l.txt
done

Split chronologically into 80% train, 10% dev, 10% test (reformat.py sorts by date).

for f in split/*.txt; do
    t=$(wc -l <$f)
    p80=$((80*t/100))
    p10=$((10*t/100))
    head -n $p80 $f > ${f%.txt}-train.txt
    tail -n +$((p80+1)) $f | head -n $p10 > ${f%.txt}-dev.txt
    tail -n +$((p80+p10+1)) $f > ${f%.txt}-test.txt
done

Create balanced dataset with 10 most frequent labels

mkdir sampled
head -n 10 label_stats.txt | awk '{ print $2 }' | while read l; do
    shuf split/${l}-train.txt | head -n 10000 > sampled/${l}-train.txt
    for t in dev test; do
        shuf split/${l}-${t}.txt | head -n 1000 > sampled/${l}-${t}.txt
    done
done
for t in train dev test; do
    cat sampled/*-${t}.txt | shuf > data/ylilauta-${t}.txt
done

Create subsets of training data

for s in 3162 1000 316 100; do
    mkdir sampled-${s}
    head -n 10 label_stats.txt | awk '{ print $2 }' | while read l; do
        shuf split/${l}-train.txt | head -n $s > sampled-${s}/${l}-train.txt
    done
    cat sampled-${s}/*-train.txt | shuf > data/ylilauta-train-${s}.txt
done

Create truncated versions of data

(Truncates lines to max 256 basic tokens, affects under 2% of examples)

mkdir trunc-data for f in data/*.txt; do python3 scripts/truncate.py 256 "$f" > trunc-data/$(basename "$f") done

Create symlinks with consistent naming (for convenience)

for s in data trunc-data; do
    mkdir $s/{100,32,10,3,1}-percent
    for d in $s/{100,32,10,3,1}-percent; do
        (
            cd $d;
            ln -s ../ylilauta-dev.txt dev.txt;
            ln -s ../ylilauta-test.txt test.txt
        )
    done
    (cd $s/1-percent; ln -s ../ylilauta-train-100.txt train.txt)
    (cd $s/3-percent; ln -s ../ylilauta-train-316.txt train.txt)
    (cd $s/10-percent; ln -s ../ylilauta-train-1000.txt train.txt)
    (cd $s/32-percent; ln -s ../ylilauta-train-3162.txt train.txt)
    (cd $s/100-percent; ln -s ../ylilauta-train.txt train.txt)
done

Experiments w/fastText

Setup

export FASTTEXT=PATH_TO_FASTTEXT

Defaults (expect ~66%)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

With more epochs and subwords (~76%)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
    -minn 3 -maxn 5 -epoch 25
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

Defaults with 10% of training data (~15%)

$FASTTEXT supervised -input data/ylilauta-train-1000.txt -output ylilauta.model
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

More epochs and subwords, 10% of training data (~32%)

$FASTTEXT supervised -input data/ylilauta-train-1000.txt -output ylilauta.model\
    -minn 3 -maxn 5 -epoch 25
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

Defaults, 1% of training data (~14%)

$FASTTEXT supervised -input data/ylilauta-train-100.txt -output ylilauta.model
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

More epochs and subwords, 1% of training data (~20%)

$FASTTEXT supervised -input data/ylilauta-train-100.txt -output ylilauta.model\
    -minn 3 -maxn 5 -epoch 25
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

Experiments with pretrained word embeddings

polyglot embeddings

Download polyglot Finnish embeddings

wget http://bit.ly/19bSmJo -O polyglot-fi.pkl

Convert to word2vec format

python3 polyglot2text.py polyglot-fi.pkl > polyglot-fi.txt

fastText with polyglot embeddings, defaults otherwise (~69%)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
     -pretrainedVectors polyglot-fi.txt -dim 64
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

More epochs and subwords (~76%)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
     -pretrainedVectors polyglot-fi.txt -dim 64 -minn 3 -maxn 5 -epoch 25
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

fastText embeddings

Download Wiki embeddings

wget https://dl.fbaipublicfiles.com/fasttext/vectors-wiki/wiki.fi.vec

Use fastText Wiki embeddings, defaults otherwise (~70)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
     -pretrainedVectors wiki.fi.vec -dim 300
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

More epochs and subwords (~76)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
     -pretrainedVectors wiki.fi.vec -dim 300 \
     -minn 3 -maxn 5 -epoch 25
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

Download and unpack Wiki + CommonCrawl embeddings

wget https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.fi.300.vec.gz
gunzip cc.fi.300.vec.gz

Use fastText embeddings, defaults otherwise (~69)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
     -pretrainedVectors cc.fi.300.vec -dim 300
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

More epochs and subwords (~75)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
     -pretrainedVectors cc.fi.300.vec -dim 300 \
     -minn 3 -maxn 5 -epoch 25
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

TurkuNLP embeddings

Download

wget http://dl.turkunlp.org/finnish-embeddings/finnish_4B_parsebank_skgram.bin

Use embeddings, defaults otherwise (~67%)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
     -pretrainedVectors finnish_4B_parsebank_skgram.bin -dim 200
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt

More epochs and subwords (~75%)

$FASTTEXT supervised -input data/ylilauta-train.txt -output ylilauta.model \
     -pretrainedVectors finnish_4B_parsebank_skgram.bin -dim 200 \
     -minn 3 -maxn 5 -epoch 25
$FASTTEXT test ylilauta.model.bin data/ylilauta-dev.txt