Cornell Movie-Dialogs Corpus
Cornell University (Cristian Danescu-Niculescu-Mizil and Lillian Lee)
304,713 lines of dialogue from 617 movie scripts, tagged by character, gender, and billing.
Fictional conversations extracted from raw scripts of 617 films: 220,579 exchanges between 10,292 pairs of characters. Each utterance links to its character and film, and films carry genre, release year, and IMDb rating. Built to study how speakers mirror each other's style, it became a staple for training chatbots.
Access
- Cost
- FreeFree download from the project page and through the ConvoKit toolkit.
- License
- License not statedNot stated on the landing page
- Sign-up
- No sign-up
- How you get it
- Bulk downloadTXT
What's in it
- One row is
- One row per utterance, grouped into conversations between two characters
- Covers
- Mostly Hollywood films
- Years
- Released 2011; films span several decades
- Updated
- One-time release
- Size
- 304,713 utterances from 617 movies and 9,035 characters
Quality
Well documented and heavily cited, but scripts are drafts rather than final films, and gender is tagged for only about 3,800 characters.
- Script text may differ from what was filmed
- Gender labels cover well under half of characters
- Static since 2011
Connects to
- IMDb Non-Commercial Datasets: Join films by title and year for cast, crew, and ratings
- MovieLens: Relate how films are written to how audiences rate them
Questions it could answer
- Do female characters get fewer lines than male characters in the same films?
- Does dialogue in higher-rated films use more varied vocabulary?