Lane's T

[Rosalee Wolfe]

The Castelloe paper from Iowa gave a very nice explanation --

>From the earlier (1976) paper, it appears that Lane first uses the 20x20 confusion matrix to create a half-matrix of similarity scores which serve as input to a cluster analysis.  The cluster analysis is the basis for building both a tree model and a 2D visualization  based on "cluster distance".  The tree model is used to build a (binary) distinctive feature model of the handshapes.  Then, as one of the techniques to measure the general ordering of the features, Lane builds a 2x2 confusion matrix for each feature (stimulus/response X present/absent) at each noise level.  The T measures shown in Figure 8 (Figure 7.8 in the later paper) are derived from groups of the 2x2 matrices using the IT measure outlined in the Castelloe paper.

It's marvelous work.  It would be interesting to have access to the original confusion matrices, to see if any of today's clustering techniques would yield a different feature model.

Rosalee


  ----- Original Message ----- 
  From: Dan Parvaz 
  To: SLLING-L at ADMIN.HUMBERC.ON.CA 
  Sent: Monday, January 26, 2004 7:27 AM
  Subject: Re: Lane's T


    Any insights into how it's done?


  I only have access to the reprint of Lane's article as found in _The Signs of Language_, which seems to skip some -- okay, all -- of the math. From my reading of the text, it isn't a 2x2, but a 20x20 confusion matrix for all the hand configurations being checked. At any rate, the Miller formulae (which is what Lane used) can be found in the first couple of pages in this paper:

  www.stat.uiowa.edu/~gwoodwor/JackBootIT.pd

  It's all that good Information Theory stuff that most CS programs (including mine!) love to torture their students with :-) If you want to talk about it, feel free to email me.

  Cheers,

  Dan.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.linguistlist.org/pipermail/slling-l/attachments/20040202/02e3aeaa/attachment.htm>