A-maze

A-maze is an incremental sentence processing method; it’s a variant of the Maze task that uses automatically generated distractors.

On other pages of this site, you can learn more about the Maze task, learn how to set-up and start using A-maze in your own research, and learn about running Maze experiments online.

This is still a work in progress, so if you run into bugs or issues, feel free to raise issues on github or email me.

Citing A-maze

If you use A-maze in your work, please cite both the A-maze and the papers about whatever model(s) you used.

• A-maze paper: V. Boyce, R. Futrell, R. P. Levy. 2020. Maze Made Easy: Better and easier measurement of incremental processing difficulty. Journal of Memory and Language.
• Gulordava model (default model): K. Gulordava, P. Bojanowski, E. Grave, T. Linzen, M. Baroni. 2018. Colorless green recurrent networks dream hierarchically. Proceedings of NAACL.
• French model: A. An, P. Qian, E. Wilcox, R. P. Levy. 2019. Representation of Constituents in Neural Language Models: Coordination Phrase as a Case Study. EMNLP 2019.

Projects using A-maze

(This is currently a very short list because I’m not aware of much work using A-maze. If you send me links to pre-prints/posters/other open access work using A-maze, I’ll add it here.)

• Maze Made Easy: Better and easier measurement of incremental processing difficulty
• “Ambiguous” isn’t “underspecified”: Evidence from the Maze task and A-maze by any other name