Re: center embedding of relative clauses

Mark Johnson (Mark.Johnson@grenoble.rxrc.xerox.com)
Tue, 12 Mar 1996 17:39:59 +0100

> I personally find it interesting that two-level embedding requires
> no padding, while three or more levels require a lot of handwaving on
> the part of the person presenting the sentence. No one else finds
> this discontinuity worthy of investigation?

Lots of us do! There's a long literature on the topic, to which I
(and another poster to this topic, Henry Thompson) have contributed.

A key paper was by none other than Chomsky and Miller, who pointed out
that while purely left or purely right branching structures can be
parsed by a finite state device, centre embedding requires memory
proportional to the depth of the embedding. Since then it has been
assumed that the ``human parser'' actually has very little working
memory, and that the centre embedded sentences are hard to process
because they require more memory than is available.

A number of us have tried to construct explicit models that predict
the amount of memory (or whatever) required to parse a given
construction and which have the property that ``comprehensible''
sentences require less memory than the ``incomprehensible'' ones.

I tried letting my own model of this loose on the Penn tree bank
version of the Brown corpus, and had it pull out all of the sentences
I would have expected to be incomprehensible. There were quite a few
(around 2%, if I recall). Coordination seemed to be most prevalent in
these sentences, suggesting that maybe coordination constructions are
not parsed in the same way as other constructions. Also, the
bracketting in the tree bank is not quite what many linguists would
claim, and my measure of the memory required is very sensitive to the
precise phrase-structure configuration involved. So the results
should be taken with a reasonable quantity of salt!

Mark

---------------------------------------------------------------------
Rank Xerox Research Centre Tel: (33) 76 61 50 37
6, chemin de Maupertuis (33) 76 61 50 50
F38240 Meylan
FRANCE Fax: (33) 76 61 50 99
---------------------------------------------------------------------