Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!lll-lcc!styx!ames!ucbcad!ucbvax!gmr.COM!MEANS
From: MEANS@gmr.COM ("Linda G. Means")
Newsgroups: comp.ai.digest
Subject: re: grammar checkers
Message-ID: <8705060504.AA14594@ucbvax.Berkeley.EDU>
Date: Mon, 4-May-87 13:18:00 EDT
Article-I.D.: ucbvax.8705060504.AA14594
Posted: Mon May  4 13:18:00 1987
Date-Received: Thu, 7-May-87 06:48:01 EDT
Sender: daemon@ucbvax.BERKELEY.EDU
Distribution: world
Organization: The ARPA Internet
Lines: 94
Approved: ailist@stripe.sri.com

mom:   toaster oven, kimono   Todd Ogasawara writes in AILIST Digest v.5 #108:

  >I think that if these style checking tools are used in conjunction 
  >with the efforts of a good teacher of writing, then these style 
  >checkers are of great benefit.  It is better that children learn a 
  >few rules of writing to start with than no rules at all.  Of course, 
  >reading lots of good examples of writing and a good teacher are still 
  >necessary.

  Sure, but the problem is the bogus rules that the child is likely
to infer from the output of the style-checking program, like never
write a sentence longer than x words, or don't use passive voice,
or try not to write sentences with multiple clauses. 


  >On another level... I happened to discuss my response above with one
  >of my dissertation committee members.  His reaction?  He pulled out
  >a recent thesis proposal filled with red pencil marks (mostly
  >grammatical remarks) and said, "So what if the style checkers are
  >superficial?  Most mistakes are superficial.  Better that the style
  >checker should find these things than me."
  
  Sounds like a rather irresponsible attitude to me, given the state
of the art of automatic style checkers.  Your prof needs a graduate
student slave if he dislikes having to correct student grammar
errors.  Let's consider separately the issues of grammar correction and
stylistic advice (the two worlds partially overlap, but remain distinct
some areas).

  1.  Grammar.  As your prof points out, lots of grammar errors are 
  superficial, but your commercial grammar checker will fail to find all
  of them, correct perceived mistakes which really aren't, and give plenty
  of bad advice.  Those programs "know" less about grammar than the students
  who use them.  Any bonafide grammatical errors which can be found by the
  commercially available software could also be found by the writer if he
  were to proof his paper carefully. It grieves me to think of students 
  failing to proof their own papers because the computer can do it for them.

  2.  Style.  The analysis of writing style is not a superficial task; it is,
  in fact, a kind of expertise not found in many "literate" individuals.
  In my experience, the best way to learn to write well is to scrutinize
  your work in the company of a good writer who will think aloud with you
  while helping you to rewrite sentences.  I've successfully taught various
  people to write that way.  The second best method is a patient teacher's 
  red pen.  In both cases, your prose is being evaluated by someone who is 
  trying to understand what you are trying to communicate in your writing.
  
     You must understand that this is not the case with the computer.  It 
  probably has no way of representing the discourse as a whole; all analysis
  is performed at the sentence level with a heavy emphasis on syntax and
  with no semantic theory of style. The result?  Stylistic advice which
  is so superficial as to be useless.  Many years of research in the area of
  computational stylistics have provided evidence that although some (few)
  stylistic discriminators can be found through syntactic analysis, the
  features which contribute to textual cohesion and to a given writer's
  "stylistic fingerprint" cannot.  Researchers are still stymied by the
  problem of identifying stylistically significant features of a text.
  Yet the program advocated by Carl Kadie feigns an understanding of the
  effect that the prose will have on its reader; it generalizes from
  syntactic structure to stylistic impact.  Look at the summary generated
  at the end of the text.  The program equates active voice and short
  sentences with "directness".  I won't take the time here to argue 
  against the use of fuzzy adjectives like 'direct', 'expressive', 'fresh',
  and so on to describe prose, since the use of such imprecise language
  is a longstanding tradition in the arena of literary criticism.  I can't
  tell you exactly how to make your writing "direct", but I know that
  directness cannot always be computed empirically, which is how your
  machine computes it.  A paragraph of non sequiturs probably shouldn't
  be characterized as direct, even if all sentences are short and contain
  only active verbs.

  An aside to Ken Laws:
  
  You questioned whether the topic of automatic style checkers is appropriate
to AILIST: is it AI?  I believe it is.  The study of computational stylistics
is a difficult natural language problem with a long history.  Topics range
from authorship studies of anonymous works to trying to identify stylistic
idiosyncrasies to automatic style advisors.  In general, many theoretical issues
carry over from other areas of natural language processing, like discourse
analysis and understanding human reasoning processes.  Think of a favorite 
author.  You may sometimes recognize a sample of his writing without
even knowing who wrote it, or you may say of another writer, "Gee, his
style reminds me of X".  You may put down a book which you started reading
because the style is too "obtuse".  How specifically does a writer use the
language to produce that effect?  What characteristics of a text must we 
identify to enable a computer to make judgments about style?  Of course, 
any advances made in tackling these issues may also be of use in the area 
of text generation.

 - Linda Means
   GM Research Laboratories
   means%gmr.com@relay.cs.net