somerandomguyontheweb/en_be_mt_datasets_evaluation

Overview

This is a small dataset of English-Belarusian sentence pairs sampled from the largest parallel corpora in OPUS (100 random instances from each of the following: NLLB, HPLT, CCMatrix, CCAligned) and manually labeled for correctness by a speaker of Belarusian. The taxonomy of labels follows Kreutzer et al. 2022:

• CC: correct translation, natural sentence
• CB: correct translation, boilerplate or low quality
• CS: correct translation, short
• X: incorrect translation
• WL: wrong language
• NL: not a language

Where appropriate, the labels are accompanied by free-form comments.

Data sampling

In Unix shell, execute:

sample_sentence_pairs () {
    mkdir -p $1
    cd $1
    wget https://object.pouta.csc.fi/OPUS-$1/$2/moses/be-en.txt.zip
    unzip be-en.txt.zip
    paste $1.be-en.en $1.be-en.be | shuf -n 100 > $1.be-en.sample100.txt
    ls | grep -v sample100 | xargs rm
    cd ..
}

sample_sentence_pairs NLLB v1
sample_sentence_pairs HPLT v2
sample_sentence_pairs CCMatrix v1
sample_sentence_pairs CCAligned v1

mv */*.txt .
rm -r NLLB HPLT CCMatrix CCAligned

Then in Python:

import csv

def to_csv(filename):
    with open(filename) as f:
        data = [line.strip().split("\t") for line in f]
    assert all(len(x) == 2 for x in data)
    with open("processed_%s.csv" % filename, "w") as f:
        csv_writer = csv.writer(f)
        csv_writer.writerow(["en", "be"])
        csv_writer.writerows(data)

to_csv("NLLB.be-en.sample100.txt")
to_csv("HPLT.be-en.sample100.txt")
to_csv("CCMatrix.be-en.sample100.txt")
to_csv("CCAligned.be-en.sample100.txt")

Labeling results

Dataset	CC	CB	CS	X	WL	NL
NLLB	17			73	10
HPLT	41	35	6	17	1
CCMatrix	7	1		92
CCAligned	31	38	8	22	1