Delivered-To: jvergne@info.unicaen.fr
Delivered-To: Jacques.Vergne@info.unicaen.fr
Organization: DFKI Saarbruecken GmbH, D 66123 Saarbruecken
Date: Mon, 27 Mar 2000 18:02:08 +0200 (MET DST)
From: COLING 2000 Programme Committee <pc@coling.org>
To: Jacques.Vergne@info.unicaen.fr
Subject: COLING 2000 Paper Number 0963
Status:   



I apologize for keeping you waiting for news about your Coling paper beyond
the deadline I had announced. I am very sorry that the committee did not
accept your paper.  We had a large number of excellent submissions and the
task of choosing among them was very difficult.


*The following are the comments that our anonymous reviewers made on your
paper. I hope you will find them useful.

***
The paper claims to present a linear algorithm for linking
non-recursive chunks in a shallow analysis system. There are three
major problems with the paper: the algorithm is very poorly described,
it seems not to be linear anyway, and there is no proper description
of evaluation results

In more detail:

section 1.1, para 3: there is a point to be made here, but
you do not succeed in making it. For phrase structure parsing,
worst-case complexity results refer to construction of the parse
forest (as long as extraction of a parse from it is efficient). When
integrated into a practical system parsers of this type only return a
fixed number of top-ranked parses so exponential behavior is avoided

section 2: how many 'virtual chunks' may exist at any one time? If
only one, how do you link inside embedded clauses? If more than one,
you have something like a stack, and then the complexity must be
proportional to the maximum stack depth (which of course is unbounded)

section 3.1, para 2: how do the decisions made in chunk linking feed
back into tagging to make it more accurate? In section 2 you seem to
describe a linear tag-chunk-link procedure

section 3.1: no evalution results given; just quoting the place you
attained in the GRACE contest is not sufficient

section 3.1, last para: the parsing rate of 0.3 seconds per word is
hardly good enough for real-time processing. This is way slower than
several current phrase structure-based full parsing systems

conclusion: again you quote the linear time claim, but the actual
parsing rate you get will rule out many internet newsfeed processing
applications

***
I could not find a precise specification of the algorithm in the paper. There are just some examples showing how the algorithm behaves for particular syntactic structures. Because of this,I could not judge if your claim on the computational cmplexity is correct. If there is a finite upper bound for the number of chunks that the algorithm looks at at each step, it is evident that the complexity is linear. But there are no descriptions suggesting that. If the upper bound is just one as your examples suggest, you should justify the point that only one chunk is enough to obtain precise links between chunks. Yet another point is that while you claim that your system does not have a "formal grammar", your rules look like yet another formal grammar. 

In addition, I think there are several linear "linking" algorithms proposed in dependency grammar community. In addition, your framework seems to have common features with deterministic parsers for CFGs. You should compare your algorithm with these works and claim that your algorithm is new in such and such respects. 

***
Though the paper gives an idea of parsing sentences based on
chunks, the presented algorithm is not precise enough to follow
the details.  The most important issue in parsing is the way to
handle ambiguities, which is not described at all in the paper.
The most essential part of the algorithm is how the only one
chunk is selected after extending a link from another chunk.
The paper simply leave this important part unstated.

The method should be compared with other related methods,
such as Finite-state cascade by Abney or Link grammars.


Thank you for giving us the opportunity to consider this paper for the
conference. I still hope we may see you in Saarbruecken!

--Martin Kay