BSD 4_2 development

[unix-history] / usr / doc / diction / rm1

.EQ
delim $$
.EN
.NH 1
Introduction
.PP
Computers have become important
in the document preparation process, with programs
to check for spelling errors and to format documents.
As the amount of text stored on line increases, it becomes
feasible and attractive to study writing
style and to attempt to help the writer in producing readable
documents.
The system of writing tools described here is a first step toward such help.
The system includes programs and a data base to
analyze writing style at the word and sentence level.
We use the term ``style'' in this paper to describe the
results of a writer's particular choices among individual words and
sentence forms.
Although many judgements of style are subjective,
particularly those of word choice,
there are some objective measures that experts
agree lead to good style.
Three programs have been written to measure some of
the objectively definable characteristics of writing style
and to identify some commonly misused or unnecessary phrases.
Although a document that conforms to the stylistic rules
is not guaranteed to be coherent and readable, one that
violates all of the rules is likely to be
difficult or tedious to read.
The program STYLE calculates readability, sentence length variability,
sentence type, word usage and sentence openers at a rate of about 400 words per second
on a PDP11/70 running the
.UX
Operating System.
It assumes that the sentences are well-formed, i. e. that
each sentence has a verb and that the subject and verb agree in number.
DICTION identifies phrases that are either bad usage or unnecessarily wordy.
EXPLAIN acts as a thesaurus for the phrases found by DICTION.
Sections 2, 3, and 4 describe the programs; Section 5 gives the results
on a cross-section of technical documents; Section 6 discusses
accuracy and problems; Section 7 gives implementation details.
.NH 1
STYLE
.PP
The program STYLE reads a document and prints a summary of
readability indices, sentence length and type, word usage,
and sentence openers.
It may also be used to locate all sentences in a document
longer than a given length, of readability index higher than a given
number, those containing a passive verb, or those beginning with an expletive.
STYLE
is based on the system for finding English word classes or parts of speech, PARTS [1].
PARTS is a set of programs that uses a small dictionary (about 350 words)
and suffix rules to partially assign word classes to
English text.
It then uses experimentally derived rules of word order to assign
word classes to all words in the text with an accuracy of about 95%.
Because PARTS uses only a small dictionary and general rules, it works
on text about any subject, from physics to psychology.
Style measures have been built into the output phase
of the programs that make up PARTS.
Some of the measures are simple counters of the word classes
found by PARTS; many are more complicated.
For example, the verb count is the total number of verb phrases.
This includes phrases like:
.DS
has been going
was only going
to go
.DE
each of which each counts as one verb.
Figure 1 shows the output of STYLE run on a paper by Kernighan and Mashey
about the
.UX
programming environment [2].
.KF
.sp 2
.TS
box;
l1l.
programming environment
readability grades:
	(Kincaid) 12.3  (auto) 12.8  (Coleman-Liau) 11.8  (Flesch) 13.5 (46.3)
sentence info:
	no. sent 335 no. wds 7419
	av sent leng 22.1 av word leng 4.91
	no. questions 0 no. imperatives 0
	no. nonfunc wds 4362  58.8%   av leng 6.38
	short sent (<17) 35% (118) long sent (>32)  16% (55)
	longest sent 82 wds at sent 174; shortest sent 1 wds at sent 117
sentence types:
	simple  34% (114) complex  32% (108)
	compound  12% (41) compound-complex  21% (72)
word usage:
	verb types as % of total verbs
	tobe  45% (373) aux  16% (133) inf  14% (114)
	passives as % of non-inf verbs  20% (144)
	types as % of total
	prep 10.8% (804) conj 3.5% (262) adv 4.8% (354)
	noun 26.7% (1983) adj 18.7% (1388) pron 5.3% (393)
	nominalizations   2 % (155)
sentence beginnings:
	subject opener: noun (63) pron (43) pos (0) adj (58) art (62) tot  67%
	prep  12% (39) adv   9% (31) 
	verb   0% (1)  sub_conj   6% (20) conj   1% (5)
	expletives   4% (13)
.TE
.sp
.ce
Figure 1
.sp
.KE
As the example shows, STYLE output is in five parts.
After a brief discussion of sentences, we will describe the parts in order.
.NH 2
What is a sentence?
.PP
Readers of documents have little
trouble deciding where the sentences end.
People don't even have to stop and think about uses of the
character ``.'' in constructions like
1.25, A. J. Jones, Ph.D., i. e., or etc. .
When a computer reads a document,
finding the end of sentences is not as easy.
First we must throw away the printer's marks and formatting
commands that litter the text in computer form.
Then STYLE
defines a sentence
as a string of words ending in one of:
.DS
 . ! ? /.
.DE
The end marker ``/.'' may be used to indicate an imperative sentence.
Imperative sentences that are not so marked are not identified as imperative.
STYLE properly handles numbers with embedded decimal points and commas,
strings of letters and numbers with embedded decimal points used for
naming computer file names, and
the common
abbreviations listed in Appendix 1.
Numbers that end sentences, like the preceding sentence, cause
a sentence break if the next word begins with a capital letter.
Initials only cause a sentence break if the next word begins with
a capital and is found in the dictionary of function words used by PARTS.
So the string
.DS
J. D. JONES
.DE
does not cause a break, but the string
.DS
 ... system H.  The ...
.DE
does.
With these rules most sentences are broken at the proper place,
although occasionally
either two sentences are called one or a fragment is called
a sentence.
More on this later.
.NH 2
Readability Grades
.PP
The first section of STYLE output consists of four readability indices.
As Klare points out in [3] readability indices may be used to
estimate the reading skills needed by the reader to understand a document.
The readability indices reported by STYLE are based on
measures of sentence and word lengths.
Although the indices
may not measure whether the document is coherent
and well organized,
experience has shown that high indices seem to be indicators of stylistic
difficulty.
Documents with short sentences and short words have low scores;
those with long sentences and many polysyllabic words have high scores.
The 4 formulae reported are Kincaid Formula [4], Automated Readability Index [5],
Coleman-Liau Formula [6]
and a normalized version of Flesch Reading Ease Score [7].
The formulae differ because they  were experimentally derived using different texts
and subject groups.
We will discuss each of the formulae briefly; for a more
detailed discussion the reader should see [3].
.PP
The Kincaid Formula, given by:
.EQ
Reading_Grade = 11.8 * syl_per_wd + .39 * wds_per_sent - 15.59
.EN
.br
was based on Navy training manuals that ranged in difficulty
from 5.5 to 16.3 in reading grade level.
The score reported by this formula tends to be in the mid-range of the
4 scores.
Because it is based on adult training manuals rather than
school book text, this formula is probably the best
one to apply to technical documents.
.PP
The Automated Readability Index (ARI), based on text from
grades 0 to 7, was derived to be easy to automate.
The formula is:
.EQ
Reading_Grade = 4.71 * let_per_wd + .5 * wds_per_sent - 21.43
.EN
.br
ARI tends to produce scores that are higher than Kincaid and
Coleman-Liau but are usually slightly lower than Flesch.
.PP
The Coleman-Liau Formula, based on text ranging in
difficulty from .4 to 16.3, is:
.EQ
Reading_Grade = 5.89 * let_per_wd - .3 * sent_per_100_wds - 15.8
.EN
.br
Of the four formulae this one usually gives the lowest
grade when applied to technical documents.
.PP
The last formula, the Flesch Reading Ease Score, is based
on grade school text covering grades 3 to 12.
The formula, given by:
.EQ
Reading_Score = 206.835 - 84.6 * syl_per_wd - 1.015 * wds_per_sent
.EN
.br
is usually reported in the range 0 (very difficult) to 100 (very easy).
The score reported by STYLE is scaled to be comparable to
the other formulas,
except that the maximum grade level reported is set to 17.
The Flesch score is usually the highest of the 4 scores
on technical documents.
.PP
Coke [8] found that the Kincaid Formula is probably the best predictor for
technical documents;
both ARI and Flesch tend to overestimate
the difficulty; Coleman-Liau tend to underestimate.
On text in the range of grades 7 to 9
the four formulas tend to be about the same.
On easy text the Coleman-Liau formula is probably
preferred since it is reasonably accurate at the lower
grades and it is safer to present text that is a little too
easy than a little too hard.
.PP
If a document has particularly difficult technical content, especially if
it includes a lot of mathematics,
it is probably best to make the text very easy to read, i.e. a lower
readability index by shortening the sentences and words.
This will allow the reader to concentrate on the technical
content and not the long sentences.
The user should remember that these indices are estimators;
they should not be taken as absolute numbers.
STYLE called with ``\-r number'' will print all sentences with
an Automated Readability Index equal to or greater than ``number''.
.NH 2
Sentence length and structure
.PP
The next two sections of STYLE output deal with sentence length and structure.
Almost all books on writing style or effective writing emphasize
the importance of variety in sentence length and structure for good writing.
Ewing's first rule in discussing style in the book
.I
Writing for Results
.R
[9] is:
.DS
``Vary the sentence structure and length of your sentences.''
.DE
Leggett, Mead and Charvat break this rule into 3 in
.I
Prentice-Hall Handbook for Writers
.R
[10] as follows:
.DS
``34a. Avoid the overuse of short simple sentences.''
``34b. Avoid the overuse of long compound sentences.''
``34c. Use various sentence structures to avoid monotony and increase effectiveness.''
.DE
Although experts agree that these rules are important, not all writers
follow them.
Sample technical documents have been found with almost no
sentence length or type variability.
One document had 90% of its sentences about the same
length as the average;
another was made up almost entirely of simple sentences (80%).
.PP
The output sections labeled ``sentence info'' and ``sentence types'' give
both length and structure measures.
STYLE reports on the number and average length of both
sentences and words,
and number of questions and imperative sentences (those ending in ``/.'').
The measures of non-function words are an attempt to look at the content
words in the document.
In English
non-function words are nouns, adjectives, adverbs, and non-auxiliary verbs;
function words are prepositions, conjunctions, articles, and auxiliary
verbs.
Since most function words are short, they tend to lower the average
word length.
The average length of non-function words may be a more useful measure for comparing
word choice of different writers than the total average word length.
The percentages of short and long sentences measure sentence
length variability.
Short sentences are those at least 5 words less than the
average; long sentences are those at least 10 words longer than the average.
Last in the sentence information section is the
length and location of the longest and shortest sentences.
If the flag ``\-l number'' is used, STYLE will print all sentences
longer than ``number''.
.PP
Because of the difficulties in dealing with the many uses of commas and conjunctions
in English, sentence type definitions
vary slightly from those of standard textbooks, but still measure
the same constructional activity.
.IP 1.
A simple sentence has one verb and no dependent clause.
.IP 2.
A complex sentence has one independent
clause and one dependent clause, each with one verb.
Complex sentences are found by identifying sentences that contain either
a subordinate conjunction or a clause beginning with words like ``that''
or ``who''.
The preceding sentence has such a clause.
.IP 3.
A compound sentence has more than one verb and no dependent
clause.
Sentences joined by ``;'' are also counted as compound.
.IP 4.
A compound-complex sentence has either several dependent clauses
or one dependent clause and a compound verb in either
the dependent or independent clause.
.PP
Even using these broader definitions, simple
sentences dominate many of the technical documents that
have been tested,
but the example in Figure 1 shows variety in both sentence structure and
sentence length.
.NH 2
Word Usage
.PP
The word usage measures are an attempt to identify
some other constructional features of writing style.
There are many different ways in English to
say the same thing.
The constructions differ from one another
in the form of the words used.
The following sentences all convey approximately the
same meaning but differ in word usage:
.DS
The cxio program is used to perform all communication between the systems.
The cxio program performs all communications between the systems.
The cxio program is used to communicate between the systems.
The cxio program communicates between the systems.
All communication between the systems is performed by the cxio program.
.DE
The  distribution of the parts of speech and verb constructions
helps identify overuse of particular constructions.
Although the measures used by STYLE are crude, they do point out
problem areas.
For each category, STYLE reports a percentage and a raw count.
In addition to looking at the percentage, the user
may find it useful to compare the raw count with the number of sentences.
If, for example, the number of infinitives is almost equal to the number
of sentences, then many of the sentences in the document are constructed
like the first and third in the preceding example.
The user may want to transform some of these sentences into another form.
Some of the implications of the word usage measures are discussed below.
.IP "\fIVerbs\fR "
are measured in several different ways to
try to determine what types of verb constructions are
most frequent in the document.
Technical writing tends to contain many
passive verb constructions and other usage of the verb ``to be''.
The category of verbs labeled ``tobe'' measures both passives and sentences of
the form:
.DS
.I
subject tobe predicate
.R
.DE
In counting verbs, whole verb phrases are counted as one verb.
Verb phrases containing auxiliary verbs are counted in the category
``aux''.
The verb phrases counted here are those whose tense is not
simple present or simple past.
It might eventually be useful to do more detailed measures
of verb tense or mood.
Infinitives are listed as ``inf''.
The percentages reported for these three categories are based on
the total number of verb phrases found.
These categories are not mutually exclusive;
they cannot be added, since, for example,
``to be going'' counts as both ``tobe'' and ``inf''.
Use of these three types of verb constructions varies significantly among authors.
.sp 2
STYLE reports passive verbs as a percentage of the finite verbs in the
document.
Most style books warn against the overuse of passive verbs.
Coleman [11] has shown that sentences with
active verbs are easier to learn than those
with passive verbs.
Although the inverted object-subject order of the passive
voice seems to emphasize the object, Coleman's experiments
showed that there is little difference in retention
by word position. He also showed that the direct object of an active verb
is retained better than the subject of a passive verb.
These experiments support the advice of the style books suggesting
that writers should try to use active verbs wherever possible.
The flag ``\-p'' causes STYLE to print all sentences containing passive verbs.
.PP
.IP "\fIPronouns\fR "
add cohesiveness and connectivity to a document
by providing back-reference.
They are often a short-hand notation for something
previously mentioned, and therefore connect the sentence containing the pronoun with the
word to which the pronoun refers.
Although there are other mechanisms for such connections, documents
with no pronouns tend to be wordy and to have little connectivity.
.IP "\fIAdverbs\fR "
can provide transition between sentences and order
in time and space.
In performing these functions, adverbs, like pronouns, provide
connectivity and cohesiveness.
.IP "\fIConjunctions\fR "
provide parallelism in a document by connecting two or more
equal units.
These units may be whole sentences, verb phrases, nouns, adjectives, or
prepositional phrases.
The compound and compound-complex sentences reported under
sentence type are parallel structures.
Other uses of parallel structures are indicated by the degree that the
number of conjunctions reported under word usage exceeds the
compound sentence measures.
.IP "\fINouns and Adjectives.\fR "
A ratio of nouns to adjectives near unity may indicate the over-use of modifiers.
Some technical writers qualify every noun with one or more
adjectives.
Qualifiers in phrases like ``simple linear single-link network model''
often lend more obscurity than precision to a text.
.IP "\fINominalizations\fR "
are verbs that are changed to nouns by adding one of the suffixes
``ment'', ``ance'', ``ence'', or ``ion''.
Examples are accomplishment, admittance, adherence, and abbreviation.
When a writer transforms a nominalized sentence to a non-nominalized
sentence, she/he increases the effectiveness of the sentence in
several ways.
The noun becomes an active verb and frequently one complicated clause
becomes two shorter clauses.
For example,
.DS
Their inclusion of this provision is admission of the importance of the system.
When they included this provision, they admitted the importance of the system.
.DE
Coleman found that the transformed sentences were easier to
learn, even when the transformation produced sentences that were
slightly longer, provided the transformation broke one clause into two.
Writers who find their document contains many
nominalizations may want to transform some of the sentences 
to use active verbs.
.NH 2
Sentence openers
.PP
Another agreed upon principle of style is variety in sentence openers.
Because STYLE determines the type of sentence opener by
looking at the part of speech of the first word in the sentence,
the sentences counted under the heading ``subject opener'' may not
all really begin with the subject.
However, a large percentage of sentences in this category
still indicates lack of variety in sentence openers.
Other sentence opener measures help the user determine
if there are transitions between sentences and where
the subordination occurs.
Adverbs and conjunctions at the beginning of sentences are mechanisms for
transition between sentences.
A pronoun at the beginning shows a link to something previously mentioned
and indicates connectivity.
.PP
The location of subordination can be determined by comparing
the number of sentences that begin with a subordinator with
the number of sentences with complex clauses.
If few sentences start with subordinate conjunctions then
the subordination is embedded or at the end of the complex sentences.
For variety the writer may want to transform some sentences
to have leading subordination.
.PP
The last category of openers, expletives, is commonly
overworked in technical writing.
Expletives are the words ``it'' and ``there'', usually with the verb ``to be'',
in constructions where the subject follows the verb.
For example,
.DS
There are three streets used by the traffic.
There are too many users on this system.
.DE
This construction tends to emphasize the object rather than the
subject of the sentence.
The flag ``\-e'' will cause STYLE to print all
sentences that begin with an expletive.
.NH 1
DICTION
.PP
The program DICTION prints all sentences in a document containing
phrases that are either frequently misused or indicate wordiness.
The program, an extension of Aho's FGREP [12] string
matching program,
takes as input a file of phrases or patterns to be matched and a file
of text to be searched.
A data base of about 450 phrases has been compiled as a default
pattern file for DICTION.
Before attempting to locate phrases, the program maps
upper case letters to lower case and substitutes blanks for
punctuation.
Sentence boundaries were deemed less critical in DICTION than
in STYLE, so abbreviations and other uses of the character
``.'' are not treated specially.
DICTION brackets all pattern matches in a sentence with the characters
``['' ``]'' .
Although many of the phrases in the default data base are correct
in some contexts, in others they indicate wordiness.
Some examples of the phrases and suggested alternatives are:
.DS
.TS
cc
ll.
Phrase	Alternative
a large number of	many
arrive at a decision	decide
collect together	collect
for this reason	so
pertaining to	about
through the use of	by or with
utilize	use
with the exception of	except
.TE
.DE
Appendix 2 contains a complete list of the default file.
Some of the entries are short forms of problem phrases.
For example, the phrase ``the fact'' is found in all of the following
and is sufficient to point out the wordiness to the user:
.DS
.TS
cc
ll.
Phrase	Alternative
accounted for by the fact that	caused by
an example of this is the fact that	thus
based on the fact that	because
despite the fact that	although
due to the fact that	because
in light of the fact that	because
in view of the fact that	since
notwithstanding the fact that	although
.TE
.DE
Entries in Appendix 2 preceded by ``~'' are not matched.
See Section 7 for details on the use of ``~''.
.PP
The user may supply her/his own pattern file with the flag ``\-f patfile''.
In this case the default file will be loaded first, followed by the user file.
This mechanism allows users to suppress
patterns contained in the default file or to include their own pet peeves that are not in the default file.
The flag ``\-n'' will exclude the default file altogether.
In constructing a pattern file, blanks should be used before and after each
phrase to avoid matching substrings in words.
For example, to find all occurrences of the word ``the'', the pattern
`` the '' should be used.
The blanks cause only the word ``the'' to be matched and not the
string ``the'' in words like there, other, and therefore.
One side effect of surrounding the words with blanks is that
when two phrases occur without intervening words, only the
first will be matched.
.NH 1
EXPLAIN
.PP
The last program, EXPLAIN, is an interactive thesaurus for
phrases found by DICTION.
The user types one of the phrases bracketed by DICTION
and EXPLAIN responds with suggested substitutions for the phrase
that will improve the diction of the document.
.KF
.DS C
Table 1
Text Statistics on 20 Technical Documents
.TS
cccccc
llnnnn.
	variable	minimum	maximum	mean	standard deviation
_
Readability	Kincaid	9.5	16.9	13.3	2.2
	automated	9.0	17.4	13.3	2.5
	Cole-Liau	10.0	16.0	12.7	1.8
	Flesch	8.9	17.0	14.4	2.2
_
sentence info.	av sent length	15.5	30.3	21.6	4.0
	av word length	4.61	5.63	5.08	.29
	av nonfunction length	5.72	7.30	6.52	.45
	short sent	23%	46%	33%	5.9
	long sent	7%	20%	14%	2.9
_
sentence types	simple	31%	71%	49%	11.4
	complex	19%	50%	33%	8.3
	compound	2%	14%	7%	3.3
	compound-complex	2%	19%	10%	4.8
_
verb types	tobe	26%	64%	44.7%	10.3
	auxiliary	10%	40%	21%	8.7
	infinitives	8%	24%	15.1%	4.8
	passives	12%	50%	29%	9.3
_
word usage	prepositions	10.1%	15.0%	12.3%	1.6
	conjunction	1.8%	4.8%	3.4%	.9
	adverbs	1.2%	5.0%	3.4%	1.0
	nouns	23.6%	31.6%	27.8%	1.7
	adjectives	15.4%	27.1%	21.1%	3.4
	pronouns	1.2%	8.4%	2.5%	1.1
	nominalizations	2%	5%	3.3%	.8
_
sentence openers	prepositions	6%	19%	12%	3.4
	adverbs	0%	20%	9%	4.6
	subject	56%	85%	70%	8.0
	verbs	0%	4%	1%	1.0
	subordinating conj	1%	12%	5%	2.7
	conjunctions	0%	4%	0%	1.5
	expletives	0%	6%	2%	1.7
.TE
.DE
.KE
.NH 1
Results
.NH 2
STYLE
.PP
To get baseline statistics and check the program's accuracy,
we ran STYLE on 20 technical documents.
There were a total of 3287 sentences in the sample.
The shortest document was 67 sentences long; the longest 339 sentences.
The documents covered a wide range of subject matter, including
theoretical computing, physics, psychology, engineering, and
affirmative action.
Table 1 gives the range, median, and standard deviation of the various style measures.
As you will note most of the measurements have a fairly wide range of values
across the sample documents.
.PP
As a comparison, Table 2 gives the median results
for two different technical authors, a sample of instructional material, and a sample of the
Federalist Papers.
The two authors show similar styles, although author 2
uses somewhat shorter sentences and longer words than author 1.
Author 1 uses all types of sentences, while author 2 prefers
simple and complex sentences, using few compound or compound-complex sentences.
The other major difference in the styles of these authors is the location
of subordination.
Author 1 seems to prefer embedded or trailing subordination, while
author 2 begins many sentences with the subordinate clause.
The documents tested for both authors 1 and 2 were technical documents,
written for a technical audience.
The instructional documents, which are written for craftspeople,
vary surprisingly little from the two technical samples.
The sentences and words are a little longer,
and they contain many passive and auxiliary verbs, few adverbs, and almost
no pronouns.
The instructional documents contain many imperative sentences, so there are
many sentence with verb openers.
The sample of Federalist Papers contrasts with the other
samples in almost every way.
.KF
.DS C
Table 2
Text Statistics on Single Authors
.TS
cccccc
llnnnn.
	variable	author 1	author 2	inst.	FED
_
readability	Kincaid	11.0	10.3	10.8	16.3
	automated	11.0	10.3	11.9	17.8
	Coleman-Liau	9.3	10.1	10.2	12.3
	Flesch	10.3	10.7	10.1	15.0
_
sentence info	av sent length	22.64	19.61	22.78	31.85
	av word length	4.47	4.66	4.65	4.95
	av nonfunction length	5.64	5.92	6.04	6.87
	short sent	35%	43%	35%	40%
	long sent	18%	15%	16%	21%
_
sentence types	simple	36%	43%	40%	31%
	complex	34%	41%	37%	34%
	compound	13%	7%	4%	10%
	compound-complex	16%	8%	14%	25%
_
verb type	tobe	42%	43%	45%	37%
	auxiliary	17%	19%	32%	32%
	infinitives	17%	15%	12%	21%
	passives	20%	19%	36%	20%
_
word usage	prepositions	10.0%	10.8%	12.3%	15.9%
	conjunctions	3.2%	2.4%	3.9%	3.4%
	adverbs	5.05%	4.6%	3.5%	3.7%
	nouns	27.7%	26.5%	29.1%	24.9%
	adjectives	17.0%	19.0%	15.4%	12.4%
	pronouns	5.3%	4.3%	2.1%	6.5%
	nominalizations	1%	2%	2%	3%
_
sentence openers	prepositions	11%	14%	6%	5%
	adverbs	9%	9%	6%	4%
	subject	65%	59%	54%	66%
	verb	3%	2%	14%	2%
	subordinating conj	8%	14%	11%	3%
	conjunction	1%	0%	0%	3%
	expletives	3%	3%	0%	3%
.TE
.DE
.KE
.NH 2
DICTION
.PP
In the few weeks that DICTION has been available
to users
about 35,000 sentences have been run with about
5,000 string matches.
The authors using the program seem to make
the suggested changes about 50-75% of the time.
To date, almost 200 of the 450 strings in the default
file have been matched.
Although most of these phrases are valid and correct
in some contexts, the 50-75% change rate seems to
show that the phrases are used much more often than
concise diction warrants.
.NH 1
Accuracy
.NH 2
Sentence Identification
.PP
The correctness of the STYLE output on the 20 document sample was checked
in detail.
STYLE misidentified
129 sentence fragments as sentences
and incorrectly joined two or more sentences 75 times
in the 3287 sentence sample.
The problems were usually because of nonstandard formatting
commands, unknown abbreviations, or lists of non-sentences.
An impossibly long sentence found as the longest sentence in
the document usually is the result of a long list
of non-sentences.
.NH 2
Sentence Types
.PP
Style correctly identified sentence type on 86.5% of
the sentences in the sample.
The type distribution of the sentences was
52.5% simple, 29.9% complex, 8.5% compound and
9% compound-complex.
The program reported 49.5% simple, 31.9% complex,
8% compound and 10.4% compound-complex.
Looking at the errors on the individual
documents, the number of simple sentences was
under-reported by about 4% and the complex and compound-complex
were over-reported by 3% and 2%, respectively.
The following matrix shows the programs output
vs. the actual sentence type.
.DS C
.TS
csssss
cccccc
clnnnn.
Program Results
		simple	complex	compound	comp-complex
Actual	simple	1566	132	49	17
Sentence	complex	47	892	6	65
Type	compound	40	6	207	23
	comp-complex	0	52	5	249
.TE
.DE
.PP
The system's inability to find imperative sentences seems to
have little effect on most of the style statistics.
A document with half of its sentences imperative was run, with and
without the imperative end marker.
The results were identical except for the expected errors of not finding
verbs as sentence openers, not counting the imperative sentences,
and a slight difference (1%) in the number of nouns
and adjectives reported.
.NH 2
Word Usage
.PP
The accuracy of identifying word types reflects
that of PARTS, which is about 95% correct.
The largest source of confusion is between nouns and
adjectives.
The verb counts were checked on about 20 sentences from each
document and found to be about 98% correct.
.NH 1
Technical Details
.NH 2
Finding Sentences
.PP
The formatting commands embedded in the text increase the difficulty
of finding sentences.
Not all text in a document is in sentence form; there are headings,
tables, equations and lists, for example.
Headings like ``Finding Sentences'' above should be discarded, not
attached to the next sentence.
However, since many of the documents are formatted to be phototypeset,
and contain font changes, which usually operate on the
most important words in the document,
discarding all formatting commands is not correct.
To improve the programs' ability to find sentence boundaries, the deformatting program, DEROFF [13],
has been given some knowledge of the formatting packages used on the
.UX
operating system.
DEROFF will now do the following:
.IP 1.
Suppress all formatting macros that
are used for titles, headings, author's name, etc.
.IP 2.
Suppress the arguments to the macros for titles, headings, author's name, etc.
.IP 3.
Suppress displays, tables, footnotes and text that is centered or in no-fill mode.
.IP 4.
Substitute a place holder for equations and check
for hidden end markers.
The place holder is necessary because many typists and authors use
the equation setter to change fonts on important words.
For this reason, header files containing the definition of
the EQN delimiters must also be included as input to STYLE.
End markers are often hidden when an equation ends a sentence
and the period is typed
inside the EQN delimiters.
.IP 5.
Add a "." after lists.
If the flag \-ml is also used, all lists are suppressed.
This is a separate flag because of the variety of ways the
list macros are used.
Often, lists are sentences that should be included in the analysis.
The user must determine how lists are used in the document to be analyzed.
.PP
Both STYLE and DICTION call DEROFF before they look at the text.
The user should supply the \-ml flag if the document contains
many lists of non-sentences that should be skipped.
.NH 2
Details of DICTION
.PP
The program DICTION is based on the string matching program FGREP.
FGREP takes as input a file of patterns to be matched and a file
to be searched and outputs each line that contains
any of the patterns
with no indication of which pattern was matched.
The following changes have been added to FGREP:
.IP 1.
The basic unit that DICTION operates on is a sentence rather than a line.
Each sentence that contains one of the patterns is output.
.IP 2.
Upper case letters are mapped to lower case.
.IP 3.
Punctuation is replaced by blanks.
.IP 4
All pattern matches in the sentence are found and surrounded with
``['' ``]'' .
.IP 5.
A method for suppressing a string match has been added.
Any pattern that begins with ``~'' will not be matched.
Because the matching algorithm finds the longest
substring, the suppression of a match allows words in some
correct contexts not to be matched while allowing
the word in another context to be found.
For example, the word ``which'' is often incorrectly used
instead of ``that'' in restrictive clauses.
However, ``which'' is usually correct when preceded by a preposition
or ``,''.
The default pattern file suppresses the match
of the common prepositions or a double
blank followed by ``which'' and therefore matches only
the suspect uses.
The double blank accounts for the replaced comma.
.NH
Conclusions
.PP
A system of writing tools that measure some of the
objective characteristics of writing style has been developed.
The tools are sufficiently general that they may be applied to
documents on any subject with equal accuracy.
Although the measurements are only of the surface
structure of the text, they do point out problem areas.
In addition to helping writers produce better documents,
these programs may be useful for studying
the writing process and finding other formulae for measuring
readability.