RaiBP/openwebtext2-first-30-chunks-lang-detect-raw-output

Counting bilingual and monolingual instances

In order to count bilingual and monolingual instances, we use the following code. We count bilingual instances where there are two languages, one of them is English and the other is either German, French, Spanish, Italian, Portuguese or Dutch. All other instances fall into the "Other" category.

from datasets import load_dataset
import json
from tqdm import tqdm


#Specify the dataset name
dataset_name = "RaiBP/openwebtext2-first-30-chunks-lang-detect-raw-output"

# Load the dataset
bilingual_dataset = load_dataset(dataset_name, data_dir='bilingual')

dataset = bilingual_dataset["train"]
n_examples = len(dataset)
keys_dict = {}
for document in tqdm(dataset, total=n_examples):

    instance_labels = document["instance_labels"]
    instance_languages = document["instance_languages"]

    for languages in instance_languages:
        unique_languages = list(set(languages))
        lang_key = "-".join(sorted(unique_languages))
        if lang_key not in keys_dict.keys():
            keys_dict[lang_key] = 1
        else:
            keys_dict[lang_key] += 1

english_keys_list = []  # keys where "en" is present
non_english_keys_list = []  # keys where "en" is not present
for key in keys_dict.keys():
    key_list = key.split('-')
    if "en" in key_list:
        english_keys_list.append(key_list)
    else:
        non_english_keys_list.append(key_list)

# more than two languages, none of them English
nen_multi_count = 0
# one language, one of the following: de, fr, es, pt, it, nl
lang_mono_count = {'de': 0, 'fr': 0, 'es': 0, 'pt': 0, 'it': 0, 'nl': 0}
# one language, not one of the following: de, fr, es, pt, it, nl
other_mono_count = 0
# two languages, none of them English
nen_bi_count = 0
for key in non_english_keys_list:
    if len(key) > 2:
        nen_multi_count += keys_dict['-'.join(key)]
    elif len(key) == 2:
        nen_bi_count += keys_dict['-'.join(key)]
    elif len(key) == 1:
        nen_lang = key[0]
        if nen_lang in lang_mono_count.keys():
            lang_mono_count[nen_lang] += keys_dict[nen_lang]
        else:
            other_mono_count += keys_dict[nen_lang]

# more than two languages, at least one of them English
english_multi_count = 0
# one language, English
english_mono_count = 0
for key in english_keys_list:
    if len(key) == 1 and key[0] == 'en':
        english_mono_count += keys_dict[key[0]]
    if len(key) > 2:
        english_multi_count += keys_dict['-'.join(key)]

# two languages, one of them English, the other one not one of the following: de, fr, es, pt, it, nl
other_bi_count = 0
# two languages, one of them English, the other one of the following: de, fr, es, pt, it, nl
lang_bi_count = {'de': 0, 'fr': 0, 'es': 0, 'pt': 0, 'it': 0, 'nl': 0}
for key in english_keys_list:
    if len(key) == 2:
        nen_lang = key[0] if key[1] == 'en' else key[1]
        if nen_lang in lang_bi_count.keys():
            lang_bi_count[nen_lang] += keys_dict['-'.join(key)]
        else:
            other_bi_count += keys_dict['-'.join(key)]

# Save the counts for monolingual
counts_dict_monolingual = {"en": english_mono_count}
for lang in lang_mono_count.keys():
    counts_dict_monolingual[lang] = lang_mono_count[lang]
counts_dict_monolingual["other"] = other_mono_count

with open('monolingual_counts.json', 'w') as json_file:
    json.dump(counts_dict_monolingual, json_file)

# Save the counts for bilingual
counts_dict_bilingual = {}
for lang in lang_bi_count.keys():
    counts_dict_bilingual[lang] = lang_bi_count[lang]
counts_dict_bilingual["other"] = other_bi_count + nen_bi_count + english_multi_count + nen_multi_count

with open('bilingual_counts.json', 'w') as json_file:
    json.dump(counts_dict_bilingual, json_file)

Counting translation instances

In order to count translation instances containing English paired with German, French, Spanish, Portuguese, Italian or Dutch, we use:

from datasets import load_dataset
import json
from tqdm import tqdm

# Specify the dataset name
dataset_name = "RaiBP/openwebtext2-first-30-chunks-lang-detect-raw-output"

# Load the dataset
translation_dataset = load_dataset(dataset_name, data_dir="translation")

dataset = translation_dataset["train"]
n_examples = len(dataset)
total_instances = 0
counts_dict = {"de": 0, "fr": 0, "es": 0, "pt": 0, "it": 0, "nl": 0}
others_count = 0
instances = {}
for document in tqdm(dataset, total=n_examples):
    embedded_label = document["embedded_label"]
    primary_label = document["primary_label"]
    document_id = document["document_index"]
    instance_id = document["instance_index"]
    id = f"{document_id}-{instance_id}"
    if id not in instances.keys():
        instances[id] = [f"{embedded_label}-{primary_label}"]
    else:
        instances[id].append(f"{embedded_label}-{primary_label}")

for id, labels in instances.items():
    state = 0
    found_langs = []
    for langs in labels:
        lang_pair = langs.split("-")
        if "en" in lang_pair:
            non_english = lang_pair[0] if lang_pair[1] == "en" else lang_pair[1]
            if non_english in counts_dict.keys():
                state = 1  # found a translation with English and a language in the counts_dict
                found_langs.append(non_english)
            elif state != 1:
                state = 2  # found a translation with English and a language not in the counts_dict
        elif state != 1:
            state = 2  # found a translation without English
    if state == 1:
        majority_lang = max(set(found_langs), key=found_langs.count)
        counts_dict[majority_lang] += 1
    elif state == 2:
        others_count += 1
    else:
        print("Error: state is 0")

# Specify the file path where you want to save the JSON file
file_path = "translation_counts.json"
counts_dict["others"] = others_count

# Save the dictionary as a JSON file
with open(file_path, "w") as json_file:
    json.dump(
        counts_dict, json_file, indent=2
    )  # indent argument is optional, but it makes the file more human-readable