It is worth emphasizing again how surprising it is that only recently,
and within the Loebner competition, have levels of performance in HMC
got back to the PARRY levels of nearly thirty years ago, in spite of
all the theoretical activity since then-it has not led to much robust
conversational performance. Even now if you check out the Chatterbots
on the web they are of astonishingly low quality, far, far below that
of the Loebner winners. A recent one, Peedy, chosen at random, told a
user that:
Just type in a sentence like:
John gave Mary a book because it was her birthday
and then you can ask the corresponding questions:
Did you give Mary a book
What did John give Mary
Who gave Mary a book
and so on.
Those who look back at Schank's inference system of 1970 will see that the example comes straight from there which is yet further proof, if it were needed, of how little movement there has been in some theoretical and empirical areas. In spite of its brusque approach to a complex technology, only the Loebner competition has been able to refocus R&D on robust and realistic performance, avoiding both the traps of theory without performance, and the performance trivialities like Chatterbot just quoted.