Evaluation Techniques for Automatic Semantic 
Extraction: Comparing Syntactic and Window Based 
Approaches 
Gregory Grefenstette 
Department of Computer Science 
University of Pittsburgh 
Pittsburgh, PA 15260 
grefen@cs.pitt, edn 
Abstract 
As large on-line corpora become more prevalent, a number of attempts have been 
made to automatically extract thesaurus-like relations directly from text using knowl- 
edge poor methods. In the absence of any specific application, comparing the results 
of these attempts is difficult. Here we propose an evaluation method using gold stan- 
dards, i.e., pre-existing hand-compiled resources, as a means of comparing extraction 
techniques. Using this evaluation method, we compare two semantic extraction tech- 
niques which produce similar word lists, one using syntactic context of words , and 
the other using windows of heuristically tagged words. The two techniques are very 
similar except that in one case selective natural language processing, a partial syn- 
tactic analysis, is performed. On a 4 megabyte corpus, syntactic contexts produce 
significantly better results against the gold standards for the most characteristk: 
words in the corpus, while windows produce better results for rare words. 
1 Introduction 
As more text becomes available electronically, it is tempting to imagine the development 
of automatic filters able to screen these tremendous flows of text extracting usefill bits of 
information. In order to properly filter, it is useful to know when two words are similar 
in a corpus. Knowing this would allcviate part of the term variability problem of natural 
language discussed in Furnas et al. (1987) . Individuals will choose a variety of words 
to name the same object or operation, with little overlap between people's choices. This 
variability in naming was cited as the principal reason for large numbers of missed citations 
in a large-scale evaluation of an information retrieval system \[Blair and Maron, 1985\]. A 
proper filter must be able to access information in the text using any word of a set of 
similar words. A number of knowledge-rich \[Jacobs and Rau, 1990, Calzolari and Bindi, 
1990, Mauldin, 1991\] and knowledge-poor \[Brown et al., 1992, Hindle, 1990, Ruge, 1991, 
Grefenstette, 1992\] methods have been proposed for recognizing when words are similar. 
The knowledge-rich approaches require either a conceptual dependency representation, or 
semantic tagging of the words, while the knowledge-poor approaches require no previously 
encoded semantic information, and depend on frequency of co-occurrence of word contexts 
to determine similarity. Evaluations of results produced by the above systems are often 
been limited to visual verification by a human subject or left to the human reader. 
In this paper, we propose gold standard evaluation techniques, allowing us to ob- 
jectively evaluate and to compare two knowledge-poor approaches for extracting word 
similarity relations from large text corpora. In order to evaluate the relations extracted, 
we measure the overlap of the results of each technique against existing hand-created 
143 
repositories of semantic information such as thesauri and dictionaries. We describe below 
}low such resources can be used as evaluation tools, and apply them to two knowledge-poor 
approaches. 
One of the tested semantic extraction approaches uses selective natural language pro- 
cessing, in this case the lexical-syntactic relations that can be extracted for each word in 
a corpus by robust parsers \[Hindle, 1983, Grefenstette, 1993\]. The other approach uses a 
variation on a classic windowing technique around each word such as was used in \[Phillips, 
1985\]. Both techniques are applied to the same 4 megabyte corpus. We evaluate the re- 
sults of both techniques using our gold standard evaluations over thesauri and dictionaries 
and compare the results obtained by the syntactic based method to those obtained by the 
windowing method. The syntax-based method provides a better overlap with the manu- 
ally defined thesaurus classes for the 600 most frequently appearing words in the corpus, 
while for rare words the windowing method performs slightly better for rare words. 
2 Gold Standards Evaluation 
2.1 Thesauri 
Roger's Thesaurus is readily available via anonymous ftp 1. In it are collected more than 
30,000 unique words arranged in a shallow hierarchy under 1000 topic numbers such as 
Existence (Topic Number 1), Inexistence (2), Substantiality (3), Unsubstantiality (4), 
..., Rite (998), Canonicals (999), and Temple (1000). Although this is far from the 
total number of semantic axes of which one could think, it does provide a wide swath of 
commonly accepted associations of English language words. We would expect that any 
system claiming to extract semantics from text should find some of the relations contained 
in this resource. 
By transforming the online source of such a thesaurus, we use it as a gold standard by 
which to measure the results of different similarity extraction techniques. This measure- 
ment is done by checking whether the 'similar words' discovered by each technique are 
placed under the same heading in this thesaurus. 
In order to create this evaluation tool, we extracted a list consisting of all single-word 
entries from our thesauri with their topic number or numbers. A portion of the extracted 
Roger list in Figure 1 shows that abatement appears under two topics: Nonincrease (36) 
and Discount (813). Abbe and abbess both belong under the same topic heading 996 
(Clergy). The extracted Roger's list has 60,071 words (an average of 60 words for each 
of the 1000 topics). Of these 32,000 are unique (an average of two occurrence for each 
word). If we assume for simplicity that each word appears under exactly 2 of the 1000 
topics, and that the words are uniformly distributed, the chance that two words wl and 
w2 occur under the same topic is 
Pnoaa = 2, (2/1000), 
since wl is under 2 topic headings and since the chance that w2 is under any specific topic 
heading is 2/1000. The probability of finding two randomly chosen words together under 
the same heading, then, is 0.4%. 
Our measurement of a similarity extraction technique using this gold standard is per- 
formed as follows. 
1 For example, in March 1993 it was available via anonymous ftp at the Internet site world.std.com in 
the directory/obi/obi2/Gutenberg/etext91, as well at over 30 other sites. 
144 
Roget ' s 
entry Topic 
,,° 
abaCement 36 
abatement 813 
abatis 717 
abatjour 260 
abattis 717 
abattoir 361 
abba 166 
abbacy 995 
abbatial 995 
abbatical 995 
abbatis 717 
~bbe 996 
abbess 996 
Macquarie 
entry subheading 
,o° 
disesteem 036406 
disesteem 063701 
diseur 022701 
disfavour 003901 
disfavour 056601 
disfavour 063701 
disfeature 018212 
disfeaturement 018201 
disfigure 006804 
disfigure 018212 
disfigure 020103 
disfigured 006803 
disfigured 020102 
.°. 
Figure 1: Samples from One Word Entries in Both Thesauri 
Given a corpus, use the similarity extraction method to derive similarity judge- 
ments between the words appearing in the corpus. For each word, take the 
word appearing as most similar. Examine the human compiled thesaurus to 
see if that pair of words appears under the same topic number. If it does, 
count this as a hit. 
This procedure was followed on the 4 megabyte corpus described below to test two seman- 
tic extraction techniques, one using syntactically derived contexts to judge similarity and 
one using window-based contexts. The results of these evaluations are also given below. 
2.2 Dictionary 
We also use an online dictionary as a gold standard following a slightly different procedure. 
Many researchers have drawn on online dictionaries in attempts to do semantic discovery 
\[Sparck Jones, 1986, Vossen et aL, 1989, Wilks et ai., 1989\], whereas we use it here only 
as a tool for evaluating extraction techniques from unstructured text. We have an online 
version of Webster's 7th available, and we use it in evaluating discovered similarity pairs. 
This evaluation is based on the assumption that similar words will share some overlap in 
their dictionary definitions. In order to determine overlap, each the entire literal definition 
is broken into a list of individual words. This list of tokens contains all the words in the 
dictionary entry, including dictionary-related markings and abbreviations. In order to 
clean this list of non-information-bearing words, we automatically removed any word or 
token 
1. of fewer than 4 characters, 
2. among the most common 50 words of 4 or more letters in the Brown corpus, 
3. among the most common 50 words of 4 or more letters appearing in the definitions 
of Webster's 7th, 
145 
ad-min-is-tra-tlon n. 1. the act or process of administering 2. performance of executive 
duties :: c<MANAGEMENT> 3. the execution of public affairs as distinguished from 
policy making 4. a) a body of persons who administer b) i<cap> :: a group constituting 
the political executive in a presidential government c) a governmental agency or board 5. 
the term of office of an administrative officer, or body. 
administer, administering, administrative, affairs, agency, board, 
constituting, distinguished, duties, execution, executive, government, 
governmental, making, management, office, officer, performance, 
persons, policy, political, presidential, public, term 
Figure 2: Webster definition of "administration," and resulting definition list after filtering 
through stoplist. 
4. listed as a preposition, quantifier, or determiner in our lexicon, 
5. of 4 or more letters from a common information retrieval stoplist, 
6. among the dictionary-related set: slang, attrib, kind, word, brit, heSS, lion, ment. 
These conditions generated a list of 434 stopwords of 4 or more characters which are 
retracted from any dictionary definition, The remaining words are sorted into a list. For 
example, the list produced for the definition of the word administration is given in Figure 2. 
For simplicity no morphological analysis or any other modifications were performed on 
the tokens in these lists. 
To compare two words using these lists, the intersection of each word's filtered defi- 
nition list is performed. For example, the intersection between the lists derived from the 
dictionary entries of diamond and ruby is (precious, stone); between right and freedom it 
is (acting, condition, political, power, privilege, right). In order to use these dictionary- 
derived lists as an evaluation tool, we perform the following experiment on a corpus. 
Given a corpus, take the similarity pairs derived by the semantic 
extraction technique in order of decreasing frequency of the first 
term. Perform the intersection of their respective two dictionary 
definitions as described above. If this intersection contains 
two or more elements, count this as a hit. 
This evahlation method was also performed on the results of both semantic extraction 
techniques applied to the corpus described in the next section. 
3 Corpus 
The corpus used for the evaluating the two techniques was extracted from Grolier's En- 
cyclopedia for other experiments in semantic extraction. In order to generate a relatively 
coherent corpus, the corpus was created by extracting only those those sentences which 
contained the word Harvard or one of the thirty hyponyms found under the word insti- 
tution in WordNet 2 \[Miller et al., 1990\], viz. institution, establishment, charity, religion, 
•.., settlement• This produced a corpus of 3.9 megabytes of text. 
2 WordNet was not used itself as a gold standard since its hierarchy is very deep and its inherent notion 
of semantic classes is not as clearly defined as in Roger. 
146 
4 Semantic Extraction Techniques 
We will use these gold standard evaluation techniques to compare two techniques for 
extracting similarity lists from raw text. 
The first technique \[Grefenstette, 1992\] extracts the syntactic context of each word 
throughout the corpus. The corpus is divided into lexical units via a regular grammar, 
each lexical unit is assigned a list of context-free syntactic categories, and a normalized 
form. Then a time linear stochastic grammar similar to the one described in \[de Marcken, 
1990\] selects a most probable category for each word. A syntactic analyzer described in 
\[Grefenstette, 1993\] chunks nouns and verb phrases and create relations within chunks 
and between chunks. A noun's context becomes all the other adjectives, nouns, and verbs 
that enter into syntactic relations with it. 
As a second technique, more similar to classical knowledge-poor techniques \[Phillips, 
1985\] for judging word similarity, we do not perform syntactic disambiguation and analysis, 
but simply consider some window of words around a given word as forming the context 
of that word. We suppose that we have a lexicon, which we do, that gives all the possible 
parts of speech for a word. Each word in the corpus is looked up in this lexicon as in 
the first technique, in order to normalize the word and know its possible parts of speech 
\[Evans et al., 1991\]. A noun's context will be all the words that can be nouns, adjectives, 
or verbs within a certain window around the noun. The window that was used was all 
nouns, adjectives, or verbs on either side of the noun within ten and within the same 
sentence. 
In both cases we will compare nouns to each other, using their contexts. In the first 
case, the disambiguator determines whether a given ambiguous word is a noun or not. In 
the second case, we will simply decide that if a word can be at once a noun or verb, or a 
noun or adjective, that it is a noun. This distinction between the two techniques of using 
a cursory syntactic analysis or not allows us to evaluate what is gained by the addition of 
this processing step. 
Figure 3 below shows the types of contexts extracted by the selective syntactic tech- 
nique and by the windowing technique for a sentence from the corpus. 
Once context is extracted for each noun, the contexts are compared for similarity 
using a weighted Jaccard measure \[Grefenstette, 1993\]. In order to reduce run time for 
the similarity comparison, only those nouns appearing more than 10 times in tile corpus 
were retained. 2661 unique nouns appear 10 times or more. For the windowing technique 
33,283 unique attributes with which to judge the words are extracted. The similarity 
judging run takes 4 full days on a DEC 5000, compared to 3 and 1/2 hours for the 
similarity calculation using data from the syntactic technique, due to greatly increased 
number of attributes for each word. For each noun, we retain the noun rated as most 
similar by the Jaccard similarity measure. Figure 4 shows some examples of words found 
most similar by both techniques. 
5 Results 
The first table, in Figure 5, compares the hits produced by the two techniques over Rogel's 
and over another online thesaurus, Macquarie's, that we had available in the Laboratory 
for Computational Linguistics at Carnegie Mellon University. This table compares the re- 
sults obtained from the windowing technique described in preceding paragraphs to those 
147 
With the arrival of Europeans in 1788 , many Aboriginal societies 
, caught vithin the coils of expanding white settlement , were 
gradually destroyed . 
Contexts o/nouns extracted after syntactic analysis 
arrival european 
society catch-SUBJ 
settlement expand-DOBJ 
Some contex~ 
arrival aboriginal 
arrival coil 
arrival settlement 
european aboriginal 
european coil 
european settlement 
society european 
society coil 
society settlement 
society aboriginal society destroy-DOBJ 
coil catch-IOBJ settlement white 
extracted with 10 full-word window 
arrival society 
arrival expand 
arrival destroy 
european society 
european expand 
european destroy 
society aboriginal 
society expand 
society destroy 
arrival catch 
arrival uhite 
european arrival 
european catch 
european ehite 
society arrival 
society catch 
society white 
Figure 3: Comparison of Extracted Contexts using Syntactic and Non-Syntactic Tech- 
niques 
Corpus word 
formation 
work 
foundation 
government 
education 
religious 
university 
group 
establishment 
power 
creation 
state 
program 
law 
year 
center 
art 
form 
century 
member 
part 
Technique used 
Syntax 
creation 
school 
institution 
constitution 
training 
religion 
institution 
institution 
creation 
authority 
establishment 
law 
institution 
constitution 
century 
development 
architecture 
group 
year 
group 
center 
system 
religious 
system 
state 
public 
century 
institution 
member 
government 
government 
state 
government 
education 
public 
government 
city 
science 
life 
religious 
group 
government 
Figure 4: Sample of words found to be most similar, by the syntactic based technique, 
and by the window technique, to some frequently occurring words in the corpus 
148 
1-20 
21-40 
41-60 
61-80 
81-100 
101-200 
201-300 
301-400 
401-500 
501-600 
601-700 
701-800 
801-900 
901-1000 
1001-2000 
2001-3000 
25% 
10% 
25% 
15% 
15% 
14% 
21% 
13% 
15% 
13% 
8% 
11% 
17% 
8% 
10.2% 
7.9% 
50% 
30% 
30% 
30% 
40% 
31% 
29% 
17% 
16% 
11% 
11% 
9% 
6% 
10% 
4.9% 
2.4% 
15% 
20% 
30% 
20% 
15% 
19% 
20% 
12% 
12% 
10% 
11% 
9% 
13% 
9% 
11.8% 
7.9% 
ROGET hits 
SYNTAX WINDOW 
40% 55% 
45% 40% 
35% 55% 
30% 45% 
35% 35% 
34% 34% 
30% 29% 
18% 25% 
13% 24% 
15% 19% 
14% 20% 
9% 17% 
7%; 25% 
9% 29% 
5.3% 19.2% 
2.1% 15.2% 
50% 
60% 
70% 
05% 
55% 
55% 
34% 
29% 
26% 
16% 
14% 
17% 
12% 
12% 
6.9% 
5.2% 
Figure 5: Windowing vs Syntactic Percentage of Hits for words from most frequent to 
least 
.c 
'3 & 
2 
results over corpus using Window vs Syntactic Contexts 
R.OGET MACQUARIE WEBSTER 
RANK WINDOW SYNTAX WINDOW SYNTAX 
1-20 21-40 41-60 61-80 81-100 100 200 300 400 500 600 700 800 900 1000 >2000 
Figure 6: Comparison of hit percentage in Roger's using simple 10-word windowing tech- 
nique(clear) vs syntactic technique(black). The y-axis gives the percentage of hits for each 
group of frequency-ranked terms. 
149 
WEBSTER hits 
°:\] '3 & • 
1-20 21-40 41-60 61-80 81-100 100 200 300 400 500 600 700 800 900 1000 >2000 
Figure 7: Comparison of hits in Macquarie's using simple 10-word windowing tech- 
nique(clear) vs syntactic technique(black). The y-axis gives the percentage of hits for 
each group of frequency-ranked terms. 
"3 
MACQUARIE hits 
%20 21-40 41'60 61-80 81-100 100 200 300 400 500 600 700 800 9(X) 1000 >2000 
Figure 8: Comparison of hit percentage in Webster's using simple 10-word windowing 
technique (hashed bars) vs syntactic technique (solid bars). The y-axis gives the percent- 
age of hits for each group of frequency-ranked terms. 
150 
Roger 
First 600 
WINDOW 
HITS 
MISS 
SYNTACTIC 
HITS MISS 
48 60 
91 401 
Macquarie SYNTACTIC 
First 600 HITS MISS 
I WINDOW 
HITS 
MISS 
42 54 
103 401 
X 2=6.4 X 2= 15.3 
p < .025 p < .005 
Roger 
Last 600 
WINDOW 
HITS 
MISS 
SYNTACTIC 
HITS MISS 
2 28 
14 556 
Macquarie 
Last 600 
WINDOW 
HITS 
MISS 
SYNTACTIC 
HITS MISS 
4 40 
14 542 
X 2=4.6 X 2= 12.5 
p < .05 p < .0005 
Figure 9: X 2 results comparing Syntactic and windowing hits in man-made thesauri 
obtained from the syntactic technique, retaining only words for which similarity judge- 
ments were made by both techniques. 
It can be seen in Figure 5 that simple technique of moving a window over a large 
corpus, counting co-occurrences of words, and eliminating empty words, provides a good 
hit ratio for frequently appearing words, since about 1 out of 5 of the 100 most frequent 
words are found similar to words appearing in the same heading in a hand-built thesaurus. 
It can also be seen that the performance of the partial syntactic analysis based tech- 
nique is better for the 600 most frequently appearing nouns, which may be considered as 
the characteristic vocabulary of the corpus. The difference in performance between the 
two techniques is statistically significant (p i 0.05). The results of a X 2 test are given in 
Figure 9. Figures 6 and 7 show the same results as histograms. In these histograms it 
becomes more evident that the window co-occurrence techniques give more hits for less 
frequently occurring words, after the 600th most frequent word. One reason for this can 
be seen by examining the 900th most frequent word, employment. Since the windowing 
technique extracts up to 20 non-stopwords from either side, there are still 537 context 
words attached to this word, while the syntactically-based technique, which examines 
finer-grained contexts, only provides 32 attributes. 
Figure 8 shows the results of applying the less focused dictionary gold standard exper- 
iment to the similarities obtained from the corpus by each technique. For this experiment, 
both techniques provide about the same overlap for frequent words, and the same signifi- 
cantly stronger showing for the rare words for the windowing technique. 
151 
6 Conclusion 
In this paper wc presented a general method for comparing tile results of two similarity 
extraction techniques via gold standards. 'Fhis method can be used when no application- 
specific evaluation technique exists and provides a relative measurement of techniques 
against human-generated standard semantic resources. We showed how these gold stan- 
dards could be processed to produce a tool for measuring overlap between their contents 
and the results of a semantic extraction method. We applied these gold standard evalu- 
ations to two different semantic extraction techniques passed over the same 4 megabyte 
corpus. The syntactic-based technique produced greater overlap with the gold standards 
derived from thesauri for the characteristic vocabulary of the corpus, while the window- 
based technique provided relatively better results for rare words. 
This dichotomous result suggests that no one statistical technique is adapted to all 
ranges of frequencies of words from a corpus. Everyday experience suggests that frequently 
occurring events can be more finely analyzed than rarer ones. In the domain of corpus 
linguistics, the same reasoning can be applied. For frequent words, finer grained context 
such as that provided by even rough syntactic analysis, is rich enough to judge similarity. 
For less frequent words, reaping more though less exact information such as that given 
by windows of N words provides more information about each word. For rare words, the 
context may have to be extended beyond a window, to the paragraph, or section, or entire 
document level, as Crouch (1990) did for rarely appearing words. 
Acknowledgements. This research was performed under the auspices of the Labora- 
tory for Computational Linguistics (Carnegie Mellon University) directed by Professor David A. 
Evans. 

References 
\[Blair and Maron, 1985\] D.C. Blair and M.E. Maron. An evaluation of retrieval effective- 
ness. Communications of the ACM, 28:289-299, 1985. 
\[Brown et al., 1992\] Peter F. Brown, Vincent J. Della Pietra, Petere V. deSouza, 
Jenifer C. Lai, and Robert L. Mercer. Class-based n-gram models of natural language. 
Computational Linguistics, 18(4):467-479, 1992. 
\[Calzolari and Bindi, 1990\] Nicoletta Calzolari and Remo Bindi. Acquisition of lexical 
information from a large textual italian corpus. In Proceedings of the Thirteenth Inter- 
national Conference on Computational Linguistics, Helsinki, 1990. 
\[Crouch, 1990\] C. J. Crouch. An approach to the automatic construction of global the- 
sauri. Information Processing and Management, 26(5):629-640, 1990. 
\[de Marcken, 1990\] Carl G. de Marcken. Parsing the LOB corpus. In 28th Annual Meeting 
of the Association for Computational Linguistics, pages 243-251, Pittsburgh, PA, June 
6-9 1990. ACL. 
\[Evans et al., 1991\] David A. Evans, Steve K. Handerson, Robert G. Lefferts, and Ira A. 
Monarch. A summary of the CLARIT project. Technical Report CMU-LCL-91-2, 
Laboratory for Computational Linguistics, Carnegie-Mellon University, November 1991. 
\[Furnas et aL, 1987\] George W. Fumas, Tomas K. Landauer, L.M. Gomez, and Susan T. 
Dumais. The vocabulary problem in human-system communication. Communications 
of the ACM, 30(11):964-971, November 1987. 
\[Grefenstette, 1992\] G. Grefenstette. Sextant: Exploring unexplored contexts for semantic 
extraction from syntactic analysis. In 30th Annual Meeting of the Association for 
Computational Linguistics, Newark, Delaware, 28 June - 2 July 1992. ACL'92. 
\[Grefenstette, 1993\] Gregory Grefenstette. Extracting semantics from raw text, imple- 
mentation details. Heuristics: the Journal of Knowledge Engineering, 1993. To Appear 
in the Special Issue on Knowledge Extraction from Text, Available as TR CS92-05, 
from the University of Pittsburgh, CS Dept. 
\[Hindle, 1983\] Donald Hindle. User manual for Fidditeh. Technical Report 7590-142, 
Navall Research Laboratory, 1983. 
\[Hindle, 1990\] D. Hindle. Noun classification from predicate-argument structures. In Pro- 
ceedings of the 28th Annual Meeting of the Association for Computational Linguistics, 
pages 268-275, Pittsburgh, 1990. ACL. 
\[Jacobs and Rau, 1990\] Paul Jacobs and Lisa Rau. SCISOR.: Extracting information from 
on-line news. Communications of the ACM, 33(11):88-97, 1990. 
\[Mauldin, 1991\] M. L. Mauldin. Conceptual Information Retrieval: A case study in adap- 
tive parsing. Kluwer, Norwell, MA, 1991. 
\[Miller et al., 1990\] George A. Miller, R. Beckwith, C. Fellbaum, D. Gross, and K. J. 
Miller. Introduction to WordNet: An on-line lexical database. Journal of Lexicography, 
3(4):235-244, 1990. 
\[Phillips, 1985\] Martin Phillips. Aspects of Text Structure: An investigation of the lexical 
organization of text. Elsevier, Amsterdam, 1985. 
\[Ruge, 1991\] Gerda Ruge. Experiments on linguistically based term associations. In 
RIAO'91, pages 528-545, Barcelona, April 2-5 1991. CID, Paris. 
\[Sparck Jones, 1986\] Karen Sparck Jones. Synonymy and Semantic Classification. Ed- 
inburgh University Press, Edinburgh, 1986. PhD thesis delivered by University of 
Cambridge in 1964. 
\[Vossen et ai., 1989\] P. Vossen, W. Meijs, and M. den Broeder. Meaning and structure 
in dictionary definitions. In Bran Boguraev and Ted Briscoe, editors, Computational 
Lexicography for Natural Language Processing, pages 171-190. Longman Group UK 
Limited, London, 1989. 
\[Wilks et al., 1989\] Yorick Wilks, D. Fass, C. Guo, J. McDonald, T. Plate, and B. Slator. 
A tractable machine dictionary as a resource for computational semantics. In Bran 
Boguraev and Ted Briscoe, editors, Computational Lexicography for Natural Language 
Processing, pages 193-228. Longman Group UK Limited, London, 1989. 
