information theory in psycholinguistics
Jul. 19th, 2007 02:02 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
guslEarlier today, talked to Chris Lucas about user modeling and intelligent tutoring. He was very enthusiastic.
---
Had lunch with Roger Levy, Vera, Christian N-M and David J. The topic was information rate, and how it's optimal to keep "constant pressure/surprise": topic-first constructions and the choice of when to use passive voice seem to agree with this. We bounced around ideas about user modeling, in order to make lecturers less clueless, and custom information-presentation systems.
Vera conjuctured that in the letter-prediction task, people will choose the first letter of most probable word (instead of the most probable letter), and now Roger wants to quick web experiment, have some fun and hopefully publish a paper. He asked us to write down our emails in his notebook. LET'S DO THIS!
Then I shared with Roger my ideas about using these Cloze-type tests as a corpus-based self-assessment tool for second-language acquisition. We need to separate knowledge of grammar and domain knowledge. He liked this and seemed enthusiastic.
It seems like I should:
* find an L2 expert, to tell me which features are good indicators of proficiency (although we could figure this out automatically by seeing where native speakers tend to make errors)
* learn Ruby on Rails / AJAX / something, to implement the following tests on the web.
** prediction tasks: letter prediction, word prediction, Cloze tests (at both levels)
** implement self-paced reading
An alternative to the above tests include playing a recording, and observing how hard people are working to process it, using:
* eye-tracking
* fMRI (David's group does this)
---
Yesterday, Mark Johnson presented a generative noise model of phonetics, which makes plausible mistakes. I'd love to see this synthesized. (I'd also like to see a voice synthesizer with foreign accents)
---
Had lunch with Roger Levy, Vera, Christian N-M and David J. The topic was information rate, and how it's optimal to keep "constant pressure/surprise": topic-first constructions and the choice of when to use passive voice seem to agree with this. We bounced around ideas about user modeling, in order to make lecturers less clueless, and custom information-presentation systems.
Vera conjuctured that in the letter-prediction task, people will choose the first letter of most probable word (instead of the most probable letter), and now Roger wants to quick web experiment, have some fun and hopefully publish a paper. He asked us to write down our emails in his notebook. LET'S DO THIS!
Then I shared with Roger my ideas about using these Cloze-type tests as a corpus-based self-assessment tool for second-language acquisition. We need to separate knowledge of grammar and domain knowledge. He liked this and seemed enthusiastic.
It seems like I should:
* find an L2 expert, to tell me which features are good indicators of proficiency (although we could figure this out automatically by seeing where native speakers tend to make errors)
* learn Ruby on Rails / AJAX / something, to implement the following tests on the web.
** prediction tasks: letter prediction, word prediction, Cloze tests (at both levels)
** implement self-paced reading
An alternative to the above tests include playing a recording, and observing how hard people are working to process it, using:
* eye-tracking
* fMRI (David's group does this)
---
Yesterday, Mark Johnson presented a generative noise model of phonetics, which makes plausible mistakes. I'd love to see this synthesized. (I'd also like to see a voice synthesizer with foreign accents)
![[identity profile]](https://www.dreamwidth.org/img/silk/identity/openid.png){kind=small}
(no subject)
Date: 2007-07-19 10:27 pm (UTC)They exist. One of the LTI speech classes builds a synthesizer each year based on a framework that leaves the voice samples up to you. (Sorry I don't know any details... I'm in MT rather than speech. Maybe try Alan Black's website?) Alan Black is from Scotland, and he once demoed to us a system that was built with his voice as the base, and it did indeed sound kind of Scottish.
(no subject)
Date: 2007-07-19 10:33 pm (UTC)