DPS Fall 2000
DCS891D Research Seminar
Stephen Parshley
September 24, 2000

Homework Assignment 1:

Select a dissertation area. You might want to pursue two in
parallel. This should be at a fairly high level initially.
As you pursue the area in greater depth, you will refine the
topic. Write a brief description of the topic (less than one
page). Make the appropriate entries in your notebook. Put
the written piece on your webpage in HTML please and send me
the URL for posting.

Introduction
I am interested in computational linguistics.  Until I am aware of the topic sufficiently, I intend to pursue two possible paths.  Professor Gustavson has helped me with suggested sources for literature research.  I have recently purchased Endnote 4.0 for windows and am using that software as my central repository for all research documentation.  However, I will still use the notebook to record my thoughts related to possible topics.

Possible Dissertation Topic 1

Automated Writing Maturity Program.  Since graduate school, I have pondered the possibility of automating the process of evaluating English compositions.  Some programs have succeeded in identifying one or two basic attributes of compositions, assigning some numerical result, and interpreting the result as an indicator of writing maturity.  Unfortunately, most such programs do not even begin to reflect the overall assessment of human graders.  This problem (being unable to emulate the subjective overall assessments of human graders for English compositions) strikes me as a potential candidate for AI-related expert systems analysis (thanks to Professor Gustavson for her guiding comments in this regard).  I foresee at least two major areas that need significant research.  First, one could attempt to capture protocols from expert, “calibrated” graders.   Developing criteria that expert graders would agree might allow the possibility of tracing comments back to specific language in the composition.  To the extent that such language can be parsed and analyzed objectively for the presence, absence, or degree of an attribute, one might be able to make inroads on a writing maturity model.  In short, the aim would be to translate subjective comments into objective analyses.  Crude attempts to do this sort of automated assessment often harshly restrict the input.  For example, a model might limit assessments to essays written in response to a known prompt.  Furthermore, the students often all have similar writing skills. The essay might have key words required and length (maximum and minimum) limitations.  Within such narrowly defined compositions, one might be able to determine relative merit.  My hope is to be able to develop assessment criteria in a more general way to apply to less defined writing samples.  Though it would be beyond the scope of my topic to develop a program and apply it to compositions, I think that I might be able to make a significant contribution by recording protocols of human experts and translating them into objective criteria for expert systems programmers.

Possible Dissertation Topic 2

Developing BNF guidelines to aid Speech Recognition Programming

Professor Ron Frank spoke with me at our first fall outing this year (September 8th 2000).  After discussing my interest and areas of expertise, he opined that there might be a need for a generalized BNF guidance for those developing Speech Recognition Programs (or even natural language programs in general).  The point here is to home in on what has yet to be done.  Speech recognition, in the sense of I say it, the computer can repeat it, is well on its way.  The next hurdle will be to do something with the input.  Since spoken language varies significantly from written language in several ways, programmers might find useful some guidance on how to develop programs to decipher text. There are many hurdles in this area, but the specific portion of this topic into which I might make inroads has to do with developing sufficient restrictions on both input and interpretation that some form of usable speech recognition might go a step further to begin “speech understanding.”  Similar to topic 1 above, this area would require categorization of speech in ways that allow a program to derive meaning in context.  Obviously, a computer that can provide meaningful prompts for clarification would be superior to one that takes what it gets and proceeds without clarification.  Just as we engage in human speech communication with feedback, a program might benefit from similar activity, acting as an interlocutor, not merely a listener.  I have no idea at this point how far one might be able to proceed along this idea, but any productive effort to narrow the scope of the problem and suggest appropriate activity would be beneficial to speech recognition research.  It is only a matter of time before complete recognition of the spoken word will be essentially perfected.  We must then turn our attention toward meaning: speech acts and the effect of context on the interpretation of speech.