metadata

license: cc-by-nc-4.0
dataset_info:
  features:
    - name: id
      dtype: string
    - name: paragraph
      dtype: int64
    - name: parlink
      dtype: int64
    - name: start
      dtype: float64
    - name: end
      dtype: float64
    - name: text
      dtype: string
    - name: times
      sequence:
        sequence: float64
    - name: source
      dtype: bool
    - name: lang
      dtype: string
    - name: retour
      dtype: string
    - name: speaker
      dtype: string
    - name: gender
      dtype: string
    - name: topic
      dtype: string
    - name: affiliation
      dtype: string
    - name: delivery
      dtype: string
    - name: audio
      dtype: audio
  splits:
    - name: train
      num_bytes: 5301635384.504
      num_examples: 11164
  download_size: 4581123984
  dataset_size: 5301635384.504
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
language:
  - pl
  - en
pretty_name: Polish Interpreting Corpus
task_categories:
  - automatic-speech-recognition
  - translation

Polish Interpreting Corpus

The Polish Interpreting Corpus (PINC) is a hand-verified parallel speech corpus derived from the European Parliament recordings. The corpus was automatically pre-processed and subsequently manually verified to correct the transcription, word-level speech-to-text alignment and sentence-level interlingual alignment. The audio quality is decent and the annotation is fairly accurate.

The corpus contains a set of 520 recordings of Polish-English speeches totalling 22 hours 58 minutes 37.92 seconds. The speeches were downloaded from the European Parliament website in the form of multi-track video files. An audio tracks containing original (EN/PL) speech and a translation in the parallel (also EN/PL) translation made by an official interpreter was extracted from the video file. Official stenographs were then aligned automatically to the audio and manually corrected by a group of reserachers. Finally, each sentence was then aligned between languages followed by the alignment of specific words within those sentences. This release of the corpus doesn't contain any word-level alignments. Also, this release contains only 1:1 alignments of text chunks. Normally, sentence-level alignment requires an M:N mapping, but multiple sentences are grouped together to achieve 1:1 mapping instead.

The basic unit of this corpus is a paragraph (ie. group of sentences). Each paragraph contains the following meta-data:

id - recording identifier as used in the original corpus
paragraph - index of the paragraph in the recording
parlink - index of the paragraph of the parallel recording (source or translation)
start - start offset in the original recording of the audio (before it was segmented for this version)
end - end offset, just like above
text - ortohgraphic transcription of the audio
times - array of time offsets (start/end) for individual (space-delimited) words from the transcription above
source - if true this segment is the source, if false it is the translation
lang - language of this segment
retour - if the interpreter is working into his/her non-native language
speaker - name of the speaker (for source segments) or unique identifier of the interpreter (for translations)
gender - male of female gender of the speaker
topic - general topic of the recording
affiliation - political affiliation of the source speaker
delivery - type of speech delivery (eg. read or impormptu speech)
audio - audio segment

The main use of this corpus was to facilitate research on PL-EN interpreting, but the data should also be useful for translation and speech technology development.

PINC was created as part of the project entitled: Extreme language control: activation and inhibition as bilingual control mechanisms in conference interpreting (research grant no. 2018/30/E/HS2/00035 funded by the Polish National Science Centre, 2019-2024).

Please cite the following work when using PINC:

@article{chmiel2023lexical,
  title={Lexical frequency modulates current cognitive load, but triggers no spillover effect in interpreting},
  author={Chmiel, Agnieszka and Janikowski, Przemys{\l}aw and Kor{\v{z}}inek, Danijel and Lijewska, Agnieszka and Kajzer-Wietrzny, Marta and Jakubowski, Dariusz and Plevoets, Koen},
  journal={Perspectives},
  pages={1--19},
  year={2023},
  publisher={Taylor \& Francis}
}

For more information please contact the Principal Investigator dr hab. Agnieszka Chmiel