/TV4Dialog

Primary LanguageSRecode Template

TV4Dialog Corpus

By Leilan Zhang

TV4Dialog is a multi-turn Chinese and English dialogue corpus, which is constructed based on scripts and subtitles of 4 TV series: Castle, Friends, House, TBBT. This corpus is suited to the fields of dialogue generation, dialogue analysis and machine translation.

License

TV4Dialog is a part of contribution of our paper Automatically Annotate TV Series Subtitles for Dialogue Corpus Construction.

The data in this repository is provided under the license CC BY 2.0. Please cite the following paper if you use the data:

@inproceedings{zhang2019,
  title={Automatically Annotate TV Series Subtitles for Dialogue Corpus Construction},
  author={Leilan Zhang, Qiang Zhou},
  year={2019},
  publisher = {{APSIPA} Press},
  address={Lanzhou, Gansu, China}
}

How TV4Dialog was made

We first collected the English scripts and Chinese-English subtitles of the 4 TV series from the Internet. The scripts were then parsed to XML format with extracted elements like scenes, speakers and utterances.

Using the methods proposed in our paper, we aligned the utterances in scripts to the subtitle lines and annotated the subtitles with speaker tags. According to those annotated tags, we merged the continuous subtitle lines belong to the same speaker to a single utterance.

Details of TV4Dialog

This corpus is composed of 26w utterances both in Chinese and English. It provides both the scripts (en) and the subtitle (en & zh), basic statistics are list below:

Directory xmlScript stores the parsed scripts of the 4 TV series in XML format (in English).

Directory withSpkr stores the subtitles annotated with speaker tags and uid tags (en & zh parallel).

Directory extracted stores the merged utterances extracted from the annotated subtitles (in Chinese).

Directory rawScript stores the raw scripts of the TV series.

Acknowledgments

We would like to thank website assrt.net for the help of providing the raw subtitle files.

We also thank websites Castle, Friends, House and TBBT for their work of collecting and sharing of the raw scripts.