Since building detectors (e.g., machine learning classifiers) seems to be unique to the domain under training and are not transferrable to other domains, there does not seem to be much opportunity for generalization of methods to assess learner states. If there are methods which tend to work across domains and/or across different people, please point me in the right direction.
Assessing cognitive and affective states in the training domain is already very difficult, and you are right that attempting domain transfer is likely to make this process even more difficult. Unfortunately for those who would like to use these tools in new data right away (and arguably fortunately for those who want to work on exciting new things), this is still a very open question. You can start with the transfer learning literature as Seyed is recommending, but from experience I can tell you that there is currently no off-the-shelf solution (that I am aware of).
There are methods to do mapping from one domain to another. These methods are either worked out algebraically (less likely) or numerically remapped using other learning techniques ( hierarchical learning)
I find in my own research on how to optimize learning of a complex task by human test subjects that it can be usefull to slow down learning, to get them to automate the basic functionality, before they must integrate those functions and apply them in more difficult tasks. As if you must ensure that sufficient attention is available early in the learning process. Not aware of the literature, I must add.
Hi Joris... Thank you for your response... my group conducts research on adaptive training... intelligent decision-making by computer-based tutors to support tailored instruction to optimize learning, retention, performance, and transfer. My RG profile has some of our publications... I recommend looking at our publication "Adaptive Intelligent Tutoring System (ITS) Research in Support of the Army Learning Model—Research Outline." It is attached... You can also find more about our work at www.GIFTtutoring.org. GIFT is our tutoring architecture... it is free to download and primarily open source.
Technical Report Adaptive Intelligent Tutoring System (ITS) Research in Suppo...
Hi Bob,
Thanks, I will read some of those and hope to learn from them. I can not provide answers to your questions but see the importance of them for a clear architecture and a sense of purpose in learners and instructors alike. It is always important to have clear goals I think and also it is often inevitable to iterate design and carefully consider all feedback.
Have you looked at Information processing theory (Theory of human problem solving), by Allen Newel and Herbert A. Simon, or Protocol analysis -verbal reports as data, by K. Anders Ericsson and Herbert A. Simon. I use these theories in my PhD for studying nurses cognitive processes and thinking strtegies.
Hege, That is ambitious of you, because much of medical problem solving is generally thought of as based on pattern recognition, because few symptoms have a one to one relationship to diseases and/or actions. But, as a tip to Bob, that thinking aloud protocols can be useful, just like teach back protocols, I heartily agree with you.
My project is about clinical reasoning. I have used the think aloud mehod, concurrent verbalizations, and analyzed with three phase protocol analysis: Reffering phrase analysis, assertional analysis and script analysis. The first phase gives you concepts that your subjects are concentrating on, the next gives you statments that subjects use by combining consepts, and the last phase you get an overview an look at the cognitive processes (i.e. inform, describe, judge...) and thinking strategies as recognising a pattern og making judgments.
Hi Hege,
Raises questions, such as how reliable are these voiced opinions, and arouses interest and admiration too, such as do you really see the reasoning and that must be fun. But... not my subject... and good luck to you!
This paper might be of interest to you: http://lp.wileypub.com/HandbookPsychology/SampleChapters/Volume4.pdf
We have just begun trying to move from Rania's Psychology analysis, to building NLP learners that have meta-cognition on issues such as optimist/pessimist smart/notsmart grumpy/pleasant by using the psychology literature/existing on-line dictionaries for these concepts, and building simple rules that associate word/frequency counts (and other metrics) with this meta data. That is, we're trying to characterize the speakers' attributes based on the NLP text. I do not have any publications on this topic, yet, however.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence