arnastofnun/IGC-2022-1

THE ICELANDIC GIGAWORD CORPUS - JSONL-FORMAT

This package contains those subcorpora of the Icelandic Gigaword Corpus, version 22.10 (http://hdl.handle.net/20.500.12537/253), that have been published with an open licence (CC-BY), in a jsonl format, which is suitable for LLM training.

---

ABOUT THE ICELANDIC GIGAWORD CORPUS (IGC):

Version 22.10 can be downloaded here: http://hdl.handle.net/20.500.12537/253

The Icelandic Gigaword Corpus (IGC) contains 8 corpora, in total almost 2,4 billion words:

Open licence:

Corpus	million words
IGC-Journals	20.9
IGC-Law	53.3
IGC-News1	396.7
IGC-Parla	254.1
IGC-Social	724.0
IGC-Wiki	8.5

Restricted licence:

Corpus	million words
IGC-Books	13.8
IGC-News2	899.8

Each corpora might contain two or more subcorpora.

Since we do not have the permission to distribute the data from Twitter (part of IGC-Social) users that download IGC have to fetch the original data themselves and then use special scripts to insert the text into the TEI-files. Due to these complications, we do not include Twitter in this package.

For further information please refer to https://igc.arnastofnun.is.

---

LICENSE:

The corpora contained in this package are published with a CC-BY license (https://creativecommons.org/licenses/by/4.0/).

---

THE HUGGINGFACE DATASET:

Each subcorpora has been converted to one jsonl-file with the "Icelandic Gigaword Corpus JSONL Converter" (http://hdl.handle.net/20.500.12537/332). Each jsonl-file belongs to one subset (configuration) in the dataset so it's possible to load each subcorpora individually.

Each line in the JSONL file contains one news article, parliamentary session etc. The information and the format of a single line is the following:

  {
      "document": "all text of the file, with paragraph splits shown as '\n\n'", 
      "uuid": "a randomly generated ID for the json object", 
      "metadata": 
      {
          "author": "the original file's author, if available", 
          "fetch_timestamp": "the date of the conversion", 
          "xml_id": "the ID of the original XML file", 
          "publish_timestamp": "the publishing date of the text in the original XML file", 
          "title": {"offset": None, "length": None},                                                  
               # the offset and length of the text's title
          "paragraphs": [{"offset": None, "length": None}, {"offset": None, "length": None}, ...],    
               # the offset and length of each paragraph
          "sentences": [{"offset": None, "length": None}, {"offset": None, "length": None}, ...],     
               # the offset and length of each sentence 
          "source": "the source of the original text, taken from the XML file"
      }
  }

Further information about each subcorpus is found in the file info.json where each of the six corpora (IGC-News1, IGC-Parla, IGC-Social ..) are a key in a dictionary holding information about the name and id of each subcorpora as well as its quality, domain, language and version)

  {
    IGC-Social: [
      {
        'subcorpus_name':'IGC-Social-Blog-heimur',
        'subcorpus_id':'social_blog_heimur',
        'quality':"A"
        'domain':['blog']	,
        'lang':"is"
        'version':"22.10"
      }
    ],
    [{}],
    ...
    IGC-Parla: [{}]
    ...
  }

Further information about the domains and how the quality of the texts was assessed is found here below.

---

USAGE:

It is possible to call each subcorpora induvidually with:

  dataset_info = load_dataset("arnastofnun/IGC-2022-1",subset)

where subset could for example be 'igc_news1_visir'

It is also possible to run a script that reads all the information about each subcorpus and loads them depending on the criteria one would set in the script:

from datasets import load_dataset, concatenate_datasets

#read dataset's configuaration 'info' that contains information about the dataset
dataset_info = load_dataset("arnastofnun/IGC-2022-1","info")
metadata = dataset_info['metadata'][0]
datasets = []
#iterate through the six corpora (Journals, Law, News1, Parla, Social, Wiki)
for corpus_name in metadata.keys():

    #ignore other corpora than IGC-News1
    if corpus_name!="IGC-News1":
        continue 
    #iterate through subcorpora
    for subcorpus in metadata[corpus_name]:

        #ignore those subcorpora that do not belong to quality category 'A'
        if subcorpus['quality']!="A":
            continue

        #load dataset
        print(subcorpus['subcorpus_id'])
        dataset=load_dataset("arnastofnun/IGC-2022-1",subcorpus['subcorpus_id'])

        #do something with the dataset ...
        #...or append to datasets to later concatenate into one dataset
        datasets.append(dataset['train'])
                
dataset_cc = concatenate_datasets(datasets)

---

CATEGORIES - DOMAINS:

We classified the 86 subcorpora into 13 domains or genres:

Adjudications
   Judgements from the three levels of jurisdiction in Iceland
   Subcorpus: IGC-Adjud
Blog
   Three online blogs
   Subcorpus: IGC-Social2
News
   Texts from general news media (online and printed)
News - local
   Texts from local news media (online and printed)
   Selected subcorpora from IGC-News1 and IGC-News2
News - radio
   Transcripts of news on the radio
   Selected subcorpora from IGC-News1
News - specialized
   Texts from media (online and printed) dedicated to specific issues (business, agriculture …)
   Selected subcorpora from IGC-News1 and IGC-News2
News - sport
   Texts from four websites that deliver sports news
   Selected subcorpora from IGC-News1 and IGC-News2
News - TV
   Transcripts of news on TV
   Selected subcorpora from IGC-News1
Online forum
   Three online forums
   Subcorpus: IGC-Social1
Parliamentary data
   Subcorpora: IGC-Parla, IGC-Law
   The Icelandic Law corpus, explanatory reports and observations extracted from bills submitted to Althingi, and parliamentary proposals and resolutions.
Published books
  Subcorpus: IGC-Books
Scientific journals
   Mainly printed journals but also a few online journals
  Subcorpus: IGC-Journals
Wikipedia
  Subcorpus: IGC-Wiki

---

CATEGORIES - QUALITY

We selected random sentences from each subcorpora (max 50.000 tokens for the bigger corpora), that were then corrected using the Byte-Level Neural Error Correction Model for Icelandic (http://hdl.handle.net/20.500.12537/255). Each sentence was also analysed with Greynir (http://hdl.handle.net/20.500.12537/269) and sentences that the tool classified as a foreign sentence were marked specially. Finally, the ratio of sentences containing errors or marked as foreign to the total amount of sentences was calculated. We divided the texts into three groups, A - C, where A has the fewest errors/foreign sentences and C the most.

As expected, texts from public data, scientific journals and news from the bigger news media (generally proofread by professionals) mostly ranked high, and texts from the online forums ranked lowest, but some texts that we had expected to rank high did not. This is due to the fact that many of the errors have nothing to do with the quality of the original text but how it was processed. Texts from Morgunblaðið, which you would expect to rank high, often had the headlines glued to the main text, which caused errors. The texts from many of the scientific journals were read with OCR which can also lead to errors. Finally, the parliamentary speeches, usually of good quality since they have been proofread, go back to the beginning of the 20th century when spelling rules were different from now.