Skip to content

Latest commit

 

History

History
61 lines (41 loc) · 4.6 KB

OUTPUT.md

File metadata and controls

61 lines (41 loc) · 4.6 KB
Raw

Bitextor output

Bitextor generates the final parallel corpora in multiple formats. These files will be placed in permanentDir folder and will have the following naming convention: {lang1}-{lang2}.{prefix}.gz, where {prefix} corresponds to a descriptor of the corresponding format.

Default

The files that will be always generated (regardless of configuration) are {lang1}-{lang2}.raw.gz and {lang1}-{lang2}.sent.gz. These files come in Moses format, i.e. tab-separated columns.

  • {lang1}-{lang2}.raw.gz: parallel corpus that contains every aligned sentences, has no deduplication and the sentences are not filtered.

    This file contains columns added by different optional modules/features: paragraph identification, deferred, Bifixer and Bicleaner. In case some of these are not enabled, the corresponding columns will be omitted. The possible fields that may appear in this file are (in this order):

    1. src_url trg_url src_text trg_text aligner_score - default columns
      • src_url and trg_url are source documents of the sentences
      • src_text and trg_text form a sentence pair in lang1 and lang2
      • aligner_score is the score given by the sentence aligner (bleualign or hunalign)
    2. src_paragraph_id trg_paragraph_id - paragraph identification data
      • initial position of the sentence in the paragraph, and initial position of the paragraph in the document
    3. src_deferred_hash trg_deferred_hash - deferred sentence hashes
    4. bifixer_hash bifixer_score - Bifixer output
      • bifixer_hash tags duplicate or near-duplicate sentences
      • bifixer_score rates quality of duplicate or near-duplicate sentences
    5. bicleaner_score - Bicleaner classifer output

    This file comes accompanied by the corresponding statistics file {lang1}-{lang2}.stats.raw, which provides information the size of the corpus in MB and in number of tokens.

  • {lang1}-{lang2}.sent.gz: parallel corpus after running all of the steps described above, plus filtering according to the specified Bicleaner threshold, adding ELRC metrics, and sorted to have the duplicate or (near-duplicate) sentences together. This file will have all of the columns from raw files, plus 3 new ones corresponding to ELRC metrics.

    1. url1 url2 sent1 sent2 aligner_score - default columns
    2. para1 para2 - paragraph identification data
    3. checksum1 checksum2 - deferred sentence checksums
    4. bifixer_hash bifixer_score - Bifixer output
    5. bicleaner_score - Bicleaner classifier output
    6. length_ratio num_tokens_src num_tokens_trg - ELRC fields
      • num_tokens_src is the number of tokens in source sentence
      • num_tokens_trg is the number of tokens in target sentence
      • length_ratio is the ratio between the two (source divided by target)

TMX

Bitextor may also generate a TMX version of the corpus, if it is enabled in config via tmx: true. The generated file is:

  • {lang1}-{lang2}.not-deduped.tmx.gz will have the same content as {lang1}-{lang2}.sent.gz, but in TMX format.

Deduplication

Bitextor may also perform the deduplication of the generated corpus (enabled via deduped: true), and generate the following files that will contain only unique sentence pairs:

  • {lang1}-{lang2}.deduped.tmx.gz: deduplicated TMX corpus, which will contain a list of URLs for the sentence pairs that were found in multiple websites.

  • {lang1}-{lang2}.deduped.txt.gz: deduplicated tab-separated corpus, with the same columns as {lang1}-{lang2}.sent.gz; in this case each sentence will only contain one source document

    Corresponding statistics file {lang1}-{lang2}.stats.deduped will be generetaed and it will have information about the size of the corpus in MB, the number of unique sentence pairs and the number of tokens on each size.

Biroamer

If Biroamer is enabled biroamer: true, Bitextor will also produce a ROAMed version of the corpus, according to Biroamer config. Biroamer will produce either of these files, depending on tmx and deduped parameters:

  • {lang1}-{lang2}.not-deduped.roamed.tmx.gz: ROAMed version of the corpus deduplicated in TMX format, that will be generated if tmx: true
  • {lang1}-{lang2}.deduped.roamed.tmx.gz: ROAMed version of the deduplicated corpus in TMX format, that will be generated if deduped: true