Berkeley-NLP/portal2

---
language:
- en
tags:
- dialogue
- multimodal
pretty_name: Portal Dialogue Corpus
size_categories:
- 10K<n<100K
---
# The Portal Dialogue Corpus

[The Portal Dialogue Corpus](https://berkeley-nlp.github.io/portal-dialogue-corpus/index.html) consists of audio, video, and game state recordings of the collaborative mode of the video game Portal 2. This folder contains the transcripts of audio conversations, along with game state information. We recorded 18 pairs of players (36 participants in total) for up to an hour each in Portal 2's cooperative mode. Players completed levels from Course 1 and Course 3 of the game, which contain 6 and 8 levels, respectively. Each file in this folder is recorded with its session number (1 through 18), course number (1 or 3), and level number (1 through 8). In particular:

* `transcripts/` - contains manually-corrected transcripts for all 153 (session, course, level) combinations. The files are structured as CSVs, with each line corresponding to an utterance from a single player. Each utterance is also automatically annotated with its communicative status, utterance type, information level, uncertainty, and discursive act. More information about these categories can be found in our paper.
* `demo_files/` - contains game state information for all 153 players, consisting of player positions and portal positions. Each row in each file corresponds to a single timestemp (tick) of the game. We also include imputed information about whether players are in each others' line of sights in the game.
* `demo_files_raw/` - Raw .dem files generated directly from Portal 2 recording. This is a proprietary format of data used by the game engine that Portal 2 is developed on, and can be parsed with various Demo File Parsers online.
For video data, please visit our [YouTube channel](https://www.youtube.com/@BerkeleyNLPSuhrGroup). For audio data, please fill out our [permissions form](https://docs.google.com/forms/d/1RgVhiGHGK2QuQhApuuCXhM7u2LmO_iig5bNDnxk0hiE/edit).