Vygotsky and Language Evolution

From WikEd

Vygotsky and LangEv

Some notes for Les Gasser's Language Evolution Seminar.

Wiki page started by Tony Hursh, feel free to contribute.

• Soviet developmental psychologist, early 20th century
• Saw learning as a social process, an interaction between a child and a knowledgeable other (Vygotsky was primarily concerned with child development, but similar ideas apply in other learning situations).
• Vygotsky's Zone of Proximal Development (ZPD)
  • The ZPD is, essentially, the area containing the activities that the student can perform only with the assistance of a knowledgeable other.
  • To a Vygotskyan, it's very important to identify the ZPD and make sure that lessons fall within it.
  • Outside the ZPD (below) -- student can do this on his/her own. Lessons are likely to be boring (waste of student and instructor time).
  • Outside the ZPD (above) -- student can't (yet) do this even with help. Lessons will be too difficult (also a waste of student and instructor time).
  • Inside the ZPD -- best use of instructor and student time.
• Vygotsky and Speech
  • Vygotsky argues that children begin by using external speech in a somewhat behavioristic way to elicit a response from caregivers, then move to "egocentric" speech (i.e., talking aloud to themselves), particularly while performing difficult tasks, and finally to internal speech, in which case complex planning occurs through speech actions. See the Tool and Symbol in Child Development chapter linked below. This was also discussed in the two papers to some degree. One way of looking at it is that the child uses egocentric speech to "give instructions to himself".
• Scenario: you're given the task of teaching an unspecified human subject to write a simple program in the Ruby programming language.
  • Nearly impossible to design this lesson for an arbitrary subject! Much easier if you have (or can acquire) more information about the subject.
    • Suppose you're told the subject is six years old?
    • Suppose you're told the subject is 22 years old?
    • Suppose you're told the 22 year old is from a small farming village, doesn't know how to read, and has only seen computers in movies?
    • Suppose you're told the 22 year old is a graduate student in Computer Science at UIUC?
    • Suppose you're allowed to give the 22 year old an extensive battery of tests to measure his/her prior knowledge of computer programming?
  • Note there is a distinction here between the difficulty of the teaching task per se and the problem of designing the teaching strategy (it would still be a very large job to teach someone who can't read how to program in Ruby, but the likelihood of a successful outcome would higher if you designed the lesson sequence taking this into account).
  • Related theorists
    • Luria -- Vygotsky's student, did a lot of work with aphasia (some possible applications to LangEv?)
    • Bruner -- Big Name in cogsci and ed psych, suggests that people code the world as hierarchical categories, and have two distinct ways of thinking about the world, narrative (telling stories, essentially) and paradigmatic (categories organized and manipulated with logic).
    • Lave and Wenger's Legitimate Peripheral Participation idea.
    • Piaget, Papert and other constructivist/constructionist ideas.

• • How could this be applied in LangEv?
    • Agent could have capabilities for egocentric/internal speech (as in the Clowes and Morse paper). This seems like an interesting approach for an autonomous agent, certainly. How well would it work in an agent society, in which the agents could both speak to themselves and to other agents? Someone needs to run some experiments here.
    • Instructor agent could attempt to acquire one or more of:
      • The current state of a specific student agent's knowledge and/or its learning history (could be probabilistic).
      • An aggregate measure of learning history over all student agents (again, could be probabilistic).
      • Effective "teaching strategies" evolved (or otherwise generated) by fellow instructor agents.
    • A simplistic (but perhaps useful) strategy would be to keep some metric (number of training iterations required, time required, etc.) of learning difficulty for each form-meaning pair, then weight "difficult" concepts to come up more often in the teaching process.
    • A more complex method could involve generating teaching rules (e.g., "teach the 'potato' form-meaning pair before teaching the 'french fry' form-meaning pair"), then exchanging these rules among instructor agents.
    • Suppose an instructor agent wants to teach the form-meaning pair "hash browns", but is able to examine the student agent's learning history and notices that it's had trouble with "french fry" and "vichyssoise". Instructor agent may be able to decide (perhaps heuristically) that the student agent might learn "hash browns" more effectively if the "potato" form-meaning pair is taught first. This would probably require more than a binary measure of "yes, I know the form-meaning pair" or "no, I don't know the form-meaning pair"; you'd also want to know how hard it was for the student agent to learn that form-meaning pair.
  • What would this cost?
    • Extra memory so the agents could carry their learning history (probably O(n) or thereabouts, where n is the number of form-meaning pairs)
    • Code for querying the learning history and current state of the student agent... adds programming difficulty (but maybe not too much).
    • Code for communicating rules/heuristics between instructor agents (again, ranging from simple "french fry is hard" rules to more complex "teach potato before french fry" rules)
  • When would this kind of strategy be useful?
    • It seems that the extra effort might be warranted if there were a large variance in the difficulty of learning form-meaning pairs, whether due to factors inherent in a specific pair itself, or the dependence of learning some pairs on prior knowledge of other pairs.

• • Further reading
    • Tool and Symbol in Child Development, from Vygotsky, L. (1930) Mind and Society
    • Interaction between Learning and Development from Vygotsky, L. (1930) Mind and Society
    • There's an excellent reading list from MIT's Open Courseware Project. It's from an MIT course called The Nature of Constructionist Learning from Spring 2003, and looks like a good overall summary of this general school of thought. Note that the Vygotsky links in the reading list are dead, but the Vygotsky links above should work.
    • The Control Systems Group looks at behavior not as controlled output but as controlled input. The agent (natural or artificial) takes actions that move its perceptions toward the perceptions corresponding to the goal (generally controlled with a cybernetic-style negative feedback loop). This is a very interesting way of looking at things. In this context, self-directed speech (internal or egocentric) could be viewed as a method of the agent controlling its own perceptions by generating artificial input data. Gary Cziko here at UIUC works with this group, and also has strong interests in evolutionary processes (he's written two books on evolution-related topics) and language, particularly what he calls autonomous technology-assisted language learning (ATALL).