PROCESSING DICTIONARY DEFINITIONS WITH PHRASAL PATTERN 
HIERARCHIES 
Hiyan Alshawi 
University of Cambridge Computer Laboratory 
Corn Exchange Street, Cambridge CB2 3QG, England* 
This paper shows how dictionary word sense definitions can be analysed by applying a hierarchy of 
phrasal patterns. An experimental system embodying this mechanism has been implemented for 
processing definitions from the Longman Dictionary of Contemporary English. A property of this 
dictionary, exploited by the system, is that it uses a restricted vocabulary in its word sense definitions. 
The structures generated by the experimental system are intended to be used for the classification of new 
word senses in terms of the senses of words in the restricted vocabulary. Examples illustrating the output 
generated are presented, and some qualitative performance results and problems that were encountered 
are discussed. The analysis process applies successively more specific phrasal analysis rules as 
determined by a hierarchy of patterns in which less specific patterns dominate more specific ones. This 
ensures that reasonable incomplete analyses of the definitions are produced when more complete 
analyses are not possible, resulting in a relatively robust analysis mechanism. Thus the work reported 
addresses two robustness problems faced by current experimental natural language processing systems: 
coping with an incomplete lexicon and with incomplete knowledge of phrasal constructions. 
INTRODUCTION 
A major factor contributing to the lack of robustness of 
experimental natural language understanding systems is 
the small number of words in the experimental semantic 
dictionaries used by these systems. For example 
"missing vocabulary" is cited as the most frequent 
cause of errors for the FRUMP system (DeJong 1979), 
a system designed to achieve a high degree of robust- 
ness. The problem does not disappear when dealing 
with limited discourse domains of the type encountered 
in database query and expert system interfaces. This is 
because of the large number of synonyms and special- 
ized words that can occur, and because of the difficulty 
of delimiting discourse domains exactly. 
A different problem faced by designers of natural 
language understanding systems is how to provide for 
graceful failure of sentence analysis. There is thus the 
need to produce reasonable incomplete interpretations 
of sentences when complete analyses are not possible. 
This situation can occur because of gaps in the gram- 
*Author's present address: SRI International, Cambridge Computer 
Science Research Centre, Millers Yard, Mill Lane, Cambridge CB2 
IRQ, England. 
matical knowledge of the system or because the system 
is faced with extragrammatical input. This paper shows 
how a possible solution to this partial analysis problem 
can be applied to the vocabulary problem in the context 
of large machine readable dictionaries. 
More specifically, we will see how word sense defi- 
nitions from the Longman Dictionary of Contemporary 
English (Procter, 1978 -- henceforth LDOCE) are proc- 
essed by a phrasal analyser that applies successively 
more specific phrasal analysis rules. The aim of this 
analysis is to provide sufficient semantic information to 
enable a system carrying out a language processing 
application to cope with occurrences of unknown 
words. 
Both the problem of coping with new words and the 
problem of robust phrasal analysis can be thought of as 
instances of a more general natural language interpreta- 
tion problem. This is the problem of coping with incom- 
plete knowledge of language use; lexical knowledge in 
the first case and knowledge of phrasal structure in the 
second. The unavoidable incompleteness of the knowl- 
edge of language use available to a language processing 
system means that trying to achieve robust natural 
language processing involves developing effective 
mechanisms for dealing with this problem. The research 
Copyright 1987 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided 
that the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To 
copy otherwise, or to republish, requires a fee and/or specific permission. 
0362-613X/87/030195-202503.00 
Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 195 
Hiyan AIshawi 
i 
Processing Dictionary Definitions with Phrasal Pattern Hiei'archies 
reported in this paper is intended to be a contribution to 
this development effort. 
The next two sections will discuss the kind of output 
that may be produced from processing dictionary defi- 
nitions and give examples of the results of processing 
LDOCE definitions produced by an implemented defi- 
nition analyser. Some problems that were encountered 
are then discussed. Later sections motivate and explain 
the basic analysis algorithm, and then describe and 
illustrate details of analysis and structure building rules. 
Finally some remarks are made about the performance 
of the current implementation and necessary further 
research. 
DEFINITION ANALYSIS 
There are various possibilities for the kind of structures 
useful for language understanding that may be derived 
from dictionary definitions. These include meaning pos- 
tulates (Carnap, 1952) expressed in some logic; con- 
straints or 'semantic formulae' based on semantic prim- 
itives (Katz and Fodor, 1963, Wilks, 1975); and 
structures carrying information enabling the classifica- 
tion of the new word sense with respect to an existing 
classification of entities in a discourse domain. The 
structures produced by the implemented definition 
analyser belong to this last type, and examples of these 
structures are given later. 
The dictionary being used in this work, LDOCE, has 
features that make it particularly suitable for definition 
analysis. Thus many LDOCE word sense entries con- 
tain additional semantic information that could be com- 
bined, or used in conjunction with, the structures pro- 
duced from processing word sense definition texts. This 
information is available as 'box codes' that give selec- 
tional restrictions, and 'subject codes' that indicate 
typical discourse domain usage of word senses (these 
codes occur in the machine-readable version of the 
dictionary, but not in the printed form). The suitability 
of LDOCE for work in computational linguistics has 
been analysed in detail by Michiels (1982). For the 
purpose of the work reported here, the most important 
property of LDOCE is the use of a restricted definition 
vocabulary of around 2000 words. Further, an impor- 
tant restriction imposed on LDOCE lexicographers is 
that only the 'central' senses of these words should 
occur in definition texts. Some ways in which the 
definitions diverge from a strict interpretation of this 
rule are discussed later. It should be remarked here that 
the LDOCE restricted definition vocabulary has more in 
common with a 'basic English' vocabulary than a set of 
semantic primitives. (A list of the words in the restricted 
definition vocabulary is given in an appendix to the 
published version (Procter, 1978) of the dictionary.) 
If the output of processing LDOCE definitions was in 
the form of meaning postulates, then the logic expres- 
sions produced would have a new symbol for the word 
sense being defined along with symbols corresponding 
to the senses of words in the definition vocabulary. 
Similarly, producing semantic primitive formulae would 
involve building new formulae by putting together for- 
mulae corresponding to the word senses of the defini- 
tion vocabulary. 
For the third possible form of output listed earlier, 
we need a (hand-coded) classification of the central 
senses of the definition vocabulary together with a 
classification of concepts in the particular domain of 
discourse in terms of these word senses. The descrip- 
tions of implementations by Bobrow and Webber 
(1980), Mark (1981), and Alshawi (1987), show how 
such a classification can be organized and used during 
text processing. The LDOCE definition for a new word 
sense is processed using the mechanism described in 
this paper in order to extract sufficient information for 
including the new word sense in such a classification. A 
natural language processing application that depended 
on a classification of concepts in the discourse domain 
should then be able to carry out its application task 
despite the occurrence of a new word in an input 
sentence. 
Extracting the information necessary for classifica- 
tion will of course include locating superordinates in the 
definitions (which define the so called "ISA" relation) 
as is done in the work reported by Amsler (1981) and 
Calzolari (1984). However, this previous work suggests 
that achieving further semantic precision in a classifica- 
tion process requires making use of other information 
present in the definition (such as modifiers and predica- 
tions). Examples of extracting this sort of information 
are presented in the next section. 
This way of dealing with unknown words in language 
processing applications still requires good solutions to 
the problem of choosing between alternative possible 
word senses (Walker and Amsler (1986) have used the 
LDOCE subject codes for this purpose) and to the 
problems involved in the classification process (see 
Schmolze and Lipkis, 1983). Nonetheless, providing a 
mechanism, as described in this paper, for extracting 
the information required by the classification process is 
a necessary first step for this approach to handling 
unknown words. 
Dictionaries vary in the level of detail provided by 
their semantic definitions, and, in general, producing a 
meaning postulate for understanding a sentence, or 
incorporating a word sense into a discourse domain 
classification, can often be done without making use of 
all the detail provided by the dictionary definition. Even 
only being able to locate the 'semantic head' (i.e. the 
main superordinate term) of a definition can be useful to 
a language processing application. This is fortunate 
since providing complete analyses of arbitrary dictio- 
nary definitions is beyond the current state of the art in 
computational linguistics. It is therefore reasonable to 
derive and make use of partial analyses of dictionary 
definitions when complete analyses are not possible. 
This is the approach taken in the implemented analysis 
system. (For certain text processing systems, for exam- 
196 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 
Hiyan Aishawi Processing Dictionary Definitions with Phrasal Pattern Hierarchies 
pie precision information retrieval systems, such an 
approach may not be acceptable.) 
ANALYSIS EXAMPLES 
I will refer to the information derived for classifying 
word senses simply as 'semantic structures'; this rather 
vague term being chosen because these structures are 
not viewed as having formal semantic status, but only as 
data structures containing information relevant to the 
classification process (or perhaps some other semantic 
process). Roughly speaking, these structures have some 
properties of a linguistic analysis of definition texts and 
some properties of a semantic definition of word sense 
concepts; their gross syntactic form is that of nested 
feature lists. 
The semantic structures are derived from various 
types of modifiers and relative clauses present in word 
sense definitions as well as the semantic head of the 
definition, if this is substantive. The syntactic category 
under which senses are grouped in the dictionary is 
important, in particular, for locating the semantic head 
of a definition and, more generally, for determining 
which analysis rules are applicable to the definition. The 
details of the analysis process are explained in a later 
section. Illustrative examples of the structures pro- 
duced from analysing the main categories of definitions 
currently handled by the implemented system -- noun, 
verb, adjective and adverb definitions -- are now given. 
Oddities in the semantic structures are often due to 
peculiarities of the current analysis grammar and output 
format, and I would not wish to argue for their correct- 
ness, especially in view of the problems discussed later. 
The following are examples of noun sense definitions 
together with the semantic structures derived from 
them. The words under which these examples occur in 
the dictionary are shown underlined. (The analysis 
system retrieves definitions from a 'lispified' version of 
the LDOCE type-setting tape, for example items pre- 
ceded by an asterisk are Lisp atoms corresponding to 
font control characters present on the type-setting tape 
(see Alshawi, Boguraev, and Briscoe, 1985).) 
(launch) 
(a large usu. motor-driven boat used for carrying 
people on rivers, lakes, harbours, etc.) 
((CLASS BOAT) (PROPERTIES (LARGE)) 
(PURPOSE 
(PREDICATION (CLASS CARRY) (OBJECT 
PEOPLE)))) 
(mug) 
(*46 BrE infml *44 a foolish person who is easily 
deceived *44 *63 see also *CA MUG'S GAME) 
((CLASS PERSON) (PROPERTIES (FOOLISH)) 
(PREDICATION (OBJECT-OF ((CLASS 
DECEIVE))))) 
(hornbeam 
(a type of small tree with hard wood, sometimes used 
in *CA HEDGE *CB *46 s) 
((CLASS TREE) (COLLECTIVE TYPE) (PROPER- 
TIES (SMALL)) 
(HAS-PART ((CLASS WOOD) (PROPERTIES 
(HARD))))) 
The semantic heads of these definitions are boat, per- 
son, and tree respectively, this being different in the last 
case from the syntactic head ("type") of the definition. 
The other information in these structures is derived 
from adjectives, prepositional phrases, and relative 
clauses. Not all the information present in the defini- 
tions is captured, for example the information conveyed 
by the phrase sometimes used in HEDGEs, in which 
HEDGE is capitalised because it is not part of the 
restricted definition vocabulary (but is defined in terms 
of this vocabulary elsewhere). 
Verb sense definitions are, in general, infinitive verb 
phrases with adverbials (often prepositional phrases) 
and additional restrictions on the semantic class of 
agents and objects. These are some examples of deriv- 
ing structures from verb sense definitions. 
(launch) 
(to send (a modern weapon or instrument) into the sky 
or space by means of scientific explosive apparatus) 
((CLASS SEND) 
(OBJECT 
((CLASS INSTRUMENT) (OTHER-CLASSES 
(WEAPON)) 
(PROPERTIES (MODERN)))) 
(ADVERBIAL ((CASE INTO) (FILLER (CLASS 
SKY)))) 
(mug) 
(to rob with violence, as in a dark street) 
((CLASS ROB) 
(ADVERBIAL ((CASE WITH) (FILLER (CLASS 
VIOLENCE))))) 
(club) 
(to beat or strike with a heavy stick (*CA CLUB *CB)) 
((CLASS STRIKE) (OTHER-CLASSES ((BEAT))) 
(ADVERBIAL 
((CASE WITH) 
(FILLER (CLASS STICK) (PROPERTIES 
(HEAVY)))))) 
Similarly, adjective sense definitions tend to have ad- 
jectival or verbal predicates as their heads, and they 
often include restrictions on the class of objects to 
which the property corresponding to the adjective can 
Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 197 
Hiyan Alshawi Processing Dictionary Definitions with Phrasal Pattern Hierarchies 
apply. The adverbial phrases used to define adverbs are 
often prepositional phrases. Examples of adjective and 
adverb definitions are the following. 
(bushy) 
((of hair) growing thickly: *46 a bushy beard / tail) 
(CLASS PROPERTY) 
(PREDICATION (CLASS GROW) (MANNER 
THICKLY)) 
(RESTRICTED-TO ((CLASS HAIR)))) 
(undomesticated) 
((of an animal) not serving man; not *CA TAME) 
((CLASS PROPERTY) 
(PREDICATION (NOT (CLASS SERVE) 
(OBJECT MAN))) 
(RESTRICTED-TO ((CLASS ANIMAL)))) 
(overland) 
(across or by land and not by sea or air) 
((MANNER ((CASE ACROSS) (FILLER ((CLASS 
LAND)))))) 
LDOCE definitions for lexicalized compound noun and 
phrasal verbs are handled in exactly the same way as 
noun and verb definitions. Two examples of structures 
generated for such definitions are given below. 
(roller coaster) 
(a kind of small railway with sharp slopes and curves, 
popular in amusement parks) 
((CLASS RAILWAY) (COLLECTIVE KIND) 
(PROPERTIES (SMALL))) 
(bring out) 
(*46 becoming rare *44 to introduce (usu. a young 
lady) into the social life of a great city *63 see also 
*CA COME OUT *CB (7)) 
((CLASS INTRODUCE) 
(OBJECT ((CLASS LADY) (PROPERTIES 
(YOUNG)))) 
(ADVERBIAL 
((CASE INTO) 
(FILLER (CLASS LIFE) (PROPERTIES 
(SOCIAL)))))) 
SOME PROBLEMS 
The current implementation is able to locate the correct 
semantic heads of dictionary definitions in most cases, 
although the examples above are untypical in the 
amount of additional information they recover from the 
definitions. Some quantitative remarks about the per- 
formance of the system are given later. This section 
briefly discusses a number of problems that were en- 
countered while testing the implemented system. 
In some respects the information conveyed by the 
output structures, being too closely tied to the surface 
definitions, only provides constraints for further seman- 
tic analysis. Perhaps the most important case of this is 
that the relationships implicit in compound nouns and 
certain prepositional phrase adverbials cannot, in gen- 
eral, be made more explicit without further interpreta- 
tion apparatus (see e.g. Alshawi, 1987) beyond that 
available to the definition analyser. The phrasal context 
can, however, sometimes allow further specification of 
relationships implicit in prepositions, for instance deri- 
vation of PURPOSE from for in cases exemplified by 
the noun sense of launch (although, of course, errors 
can result from attempting to make relationships more 
explicit in this way), The actual words appearing in the 
semantic structures are, on the other hand, further 
disambiguated than might be assumed given the high 
degree of polysemy of many of the words in the re- 
stricted vocabulary. This is because the analysis proc- 
ess identifies the syntactic category of these words and 
because of the LDOCE rule that only the most central 
senses of words from the restricted vocabulary should 
appear in definitions (but see the remarks below on 
phrasal verbs). 
The fact that definition texts are often not analysed 
completely means that information that is central to a 
definition is sometimes not taken into account, as 
illustrated by the following example. In this case the 
usual 'purpose' of nails is not recovered. 
(nail) 
(a thin piece of metal with a point at one end and a fiat 
head at the other for hammering into a piece of wood, 
usu. to fasten the wood to something else) 
((CLASS PIECE) (MATERIAL METAL) (PROPER- 
TIES (THIN)) 
(HAS-PART ((CLASS POINT)))) 
Although the base forms of all the words in the 
restricted vocabulary and simple morphological vari- 
ants of these are handled by the analysis process, there 
are many cases of derivational morphology which are 
not currently handled. Difficulties are also caused by the 
liberal use in LDOCE definitions of phrasal verbs made 
up from verbs and particles taken from the restricted 
vocabulary. The idiomatic nature of phrasal verbs 
means that the rule of using only the central senses of 
words in the definition vocabulary is violated in many 
cases in which phrasal verbs are used to implicitly 
increase the size of the defining vocabulary. An exam- 
ple is the occurrence of look after and bring up causing 
an error in the analysis of the following sense definition 
for foster: 
(foster) 
(to look after or bring up (a child or young animal) as 
one's own . . .) 
198 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 
Hiyan Aishawi Processing Dictionary Definitions with Phrasal Pattern Hierarchies 
((CLASS LOOK)). 
Another problem encountered in LDOCE entries is 
that word senses are sometimes defined in terms of 
previous senses of the same homograph. For example 
immediately after the definition of the sense of horn- 
beam given earlier, the following dependent word sense 
definition is present, for which the system produces a 
structure containing the special symbol '*previous- 
sense*'. 
(the wood of this tree) 
((CLASS WOOD) (RELATED-TO 
*PREVIOUS-SENSE*)) 
A problem related to the one just mentioned is that only 
the simplest forms of cross references to words not 
included in the definition vocabulary are handled at 
present. However, given the compositional nature of 
nested feature lists, and the fact that definitional cross 
references are intended to be non-circular in LDOCE, it 
should be feasible to use semantic structures for the 
referenced words (and previous senses as in the horn- 
beam example) in building other semantic structures. 
The use of a restricted vocabulary in LDOCE defi- 
nitions means that the lexicographers have already 
engaged in a substantial amount of semantic analysis of 
word senses that is potentially useful for automatic 
natural language processing. However, as observed by 
Michiels (1982), there is a tradeoff between the size of 
the definition vocabulary and the syntactic complexity 
of definitions. This implies that in order to take full 
advantage of the potential of LDOCE entries for lan- 
guage processing we need to pay special attention to the 
design of the definition analyser; this is the issue ad- 
dressed in the rest of this paper. 
PHRASAL ANALYSIS HIERARCHIES 
The analysis mechanism has the flavour of a pattern- 
based phrasal analyser. It was designed to overcome 
some of the more obvious difficulties of applying a 
simple pattern matching approach to robust phrasal 
analysis. In particular it was required that the mecha- 
nism should have the means to specify which compo- 
nents of a phrase are more important, and to index 
analysis rules so that the mechanism would be reason- 
ably efficient. 
Pattern matching has played an important role in 
several previous parsing systems, for example those of 
Wilks (1975), Parkinson et al (1977), Wilensky and 
Arens (1980), and Hayes and Mouradian (1981). A 
characteristic of the mechanism used here is that it 
depends on a hierarchy of phrasal analysis patterns in 
which more specific patterns are dominated by less 
specific ones. 
The basic analysis algorithm that applies the hierar- 
chy of patterns is as follows. Starting at the top of the 
hierarchy, a pattern is matched against the input defini- 
tion, If the match with this pattern succeeds then a 
match is attempted with each of its daughter patterns 
(i.e. the more specific forms of this pattern placed 
immediately below it in the hierarchy). This procedure 
is repeated recursively so that we end up with the most 
specific matches against the input definition. This pars- 
ing technique is different in kind from the more common 
approach to robust parsing in which exact grammar 
rules are tried first before being relaxed by the parser 
(see e.g. Weischedel and Black, 1980, Kwasny and 
Sondheimer, 1981, and Pulman, 1984). 
The hierarchy provides a natural solution to the 
indexing problem mentioned above since it restricts the 
application of patterns to those that are more likely to 
succeed, enabling efficient phrasal analysis. It also 
provides a solution to the problem of specifying the 
more important components of phrases since less spe- 
cific patterns tend to be concerned with more important 
components only. This ensures that reasonable incom- 
plete analyses can be produced when more detailed 
analyses are not possible. 
Each analysis rule consists of a rule identifier, a 
phrasal pattern, and a list of rule identifiers for daughter 
patterns. It is written in the following form: 
(rule identifier phrasal pattern daughter identifiers). 
The rule identifier also appears in a semantic struc- 
ture building rule. These two types of rule are kept 
separate in order to allow different kinds of output 
structures to be generated for the same analysis gram- 
mar. Building semantic structures is basically a simple 
process of fleshing out templates provided by the se- 
mantic structure building rules using variable bindings 
generated by the matching algorithm. 
The following section gives some examples of anal- 
ysis and structure building rules, explaining the notation 
in which the phrasal patterns are written. The notation 
currently provides a limited number of facilities, but it 
should be clear that these facilities could be extended in 
various ways while remaining within the the overall 
framework of applying a hierarchy of phrasal patterns 
as discussed above. 
ANALYSIS RULES 
A typical analysis rule, n-100, for noun definitions, and 
two of its descendants, n-ll0 and n-135, are shown 
below, n-ll0 is a daughter of n-100, and n-135 is a 
daughter of n-130 (not shown). More mnemonic identi- 
fiers for these rules might be "Noun-phrase", "Simple- 
NP", and "NP-with-relative" instead of n-100, n-ll0, 
and n-135 respectively. 
(n-100 (n && +0det && &0adj &noun &&) n-ll0 n-120 
n-130 n-140) 
(n-110 (n +0det +0intens &0adj &noun *0pp-mod &&)) 
(n-135 (n +0det &0adj &0noun +nounl +that-which 
*verb-pred &&)) 
Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 199 
Hiyan Alshawi Processing Dictionary Definitions with Phrasal Pattern Hierarchies 
In the phrasal pattern part of these rules, the initial 
"n" restricts the pattern to matching definitions for 
senses with lexical category "n", i.e. nouns. The other 
pattern elements match zero or more items in the input 
depending on the type of the element (indicated by its 
first one or two characters) and restrictions in terms of 
lexical features. Digits at the end of pattern elements 
simply distinguish different occurrences of elements 
with the same properties. Examples of pattern elements 
and what they can match are the following. 
for 
÷ noun 
+Odet 
&noun 
&0adj 
&& 
*0pp-mod 
*verb-pred 
the word for 
exactly one noun 
zero or one determiner 
one or more nouns 
zero or more adjectives 
an arbitrary segment of input words 
zero or one prepositional phrase 
modifier 
a segment that matches a verb 
phrase pattern 
The last element is an example of an element with 
subsidiary patterns (in this case for verb phrases) which 
use the same kinds of element as above. There are thus 
associated with this element a list of rules including 
'(passive-pred (be +vtrans))'. Here 'passive-pred' is 
just the name of the subsidiary pattern '(be +vtrans)'. 
Similarly, one subsidiary pattern of '*0pp-mod' is '(for- 
pp (+0used for +noun-verb)'. The use of elements with 
such subsidiary patterns allows for recursion and a 
more compact set of patterns, in much the same way as 
for conventional context free phrase structure gram- 
mars. 
Given this interpretation of pattern elements, it 
should be clear that the phrasal patterns of rules n-110 
and n-135 will match subsets of the set of definitions 
matched by the phrasal pattern of rule n-100. The noun 
sense definition examples given earlier for launch and 
mug matched n-110 and n-135 respectively. (There is an 
initial morphological analysis phase which discards 
items like usu. which it does not recognize.) The 
analysis algorithm outlined in the previous section en- 
sures that n-ll0 and n-135 are tried only if n-100 
succeeds. 
The structure building rule associated with an analy- 
sis rule is applied when none of the immediate descen- 
dants of the analysis rule succeed. Thus the structure 
building rule for n-100 is applied when none of n-ll0, 
n-120, n-130, or n-140 succeed, ensuring that a semantic 
structure is built according to the analysis provided by 
n-100 if no more specific version of this analysis is 
possible. The structure building rules for n-100 and 
n-135 are given below. 
(n-100 ((compound-class &noun) (properties &0adj))) 
(n-135 ((class +nounl) (noun-mods &0noun) 
(properties &0adj) (predication *verb-pred))) 
The semantic structure given earlier for the definition a 
foolish person who is easily deceived . . . (a British 
English sense of mug) was generated using the rule just 
given above. Applying the rule involves replacing the 
variables with bindings generated by the matching proc- 
ess; splicing-out substructures associated with unin- 
stantiated optional pattern elements; and recursively 
applying this process to the structure building rule 
associated with the appropriate (i.e. successfully match- 
ing) subsidiary pattern for the element '*verb-pred'. 
This last step results in building the substructure 
"(predication (object-of ((class deceive)))" using a rule 
associated with the subsidiary pattern 'passive-pred' 
that was mentioned earlier. There is also an optional 
further stage of the structure building process which 
applies transformations specified as attached proce- 
dures associated with items in the structure building 
rules, for example the item 'predication'. This phase 
gives greater freedom than would be possible by the use 
of structure building templates alone, for example it 
allows moving items (such as those indicating negation) 
'upwards' out of substructures. 
The analysis algorithm follows all paths from suc- 
cessful matches. This does not lead to inefficiency 
because it is rare for several deep, but disjoint, paths to 
be followed successfully down the pattern hierarchy, 
and because the implementation maintains a well- 
formed substring table to avoid a certain amount of 
redundant computation. Alternative semantic struc- 
tures can result from processing a definition when there 
is more than one most specific successful analysis rule. 
At present one such analysis is chosen by an over 
simplistic heuristic that basically prefers analyses ac- 
counting for more words of the input definition. 
PERFORMANCE REMARKS 
Although some of the difficulties mentioned in the 
section on problems are not easy to accommodate in the 
present system, most of the errors in identifying seman- 
tic heads of definitions were not due to these. Instead 
these errors appear to be caused mainly by failure on 
the part of the rudimentary morphological analysis 
performed, and the inadequate coverage of the present 
set of phrasal patterns. 
In order to evaluate the performance of the system, it 
was tested against a sample of 500 definitions chosen at 
random using a facility for automatic random selection 
of entries provided by the dictionary access system. 
Only a few of these definitions will have coincided with 
those processed during the development of the system 
and its analysis rules. The selection process ignored 
definitions for idioms and those with complex cross 
references, so these are not taken into account in the 
figures given below. 
The results of the test were as follows. The semantic 
200 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 
Hiyan Alshawi Processing Dictionary Definitions with Phrasal Pattern Hierarchies 
head was identified correctly for 387 definitions (77%). 
Additional information was recovered for 236 (61%) of 
these definitions, and this additional information was 
judged to be correct for 207 (88%) of these cases. 
Thus only identifying the head is much more typical 
than might be suggested by the examples given earlier 
for illustrative purposes, and in fact the present set of 
rules rarely takes into account all the words appearing 
in a definition. There are altogether some 90 phrasal 
patterns in the hierarchy, including subsidiary patterns. 
It should be emphasized that this grammar of definitions 
was written as a feasibility test, and I believe it is 
reasonable to expect that the number of patterns in the 
hierarchy could be enlarged to 400, say, before the 
problem of diminishing returns becomes a serious one. 
The current implementation of the definition analyser 
takes around half a second to access the dictionary and 
process a definition. This is the elapsed time on a lightly 
loaded GEC-63 (a 32 bit mini-computer); the definition 
analyser was implemented in Cambridge Lisp and low 
level dictionary access in the language C. Finally, 
perhaps it is worth mentioning that the development 
effort was only a few man months for each of the 
program and grammar, which, compared with other 
natural language processing systems we have developed 
recently at Cambridge, represents a relatively small 
effort. 
FURTHER RESEARCH 
The work carried out so far seems to suggest that 
dictionary definitions can be analysed with a reasonable 
degree of success using hierarchies of phrasal patterns, 
but it still remains to be demonstrated that this tech- 
nique can enable an actual natural language application 
system to cope effectively with unknown words. 
Although dictionary definitions exhibit a rich variety 
of forms, these are mostly variations on a manageably 
small number of basic forms, and it is this property of 
definitions that makes phrasal pattern hierarchies par- 
ticularly appropriate for analysing them. It seems likely 
however that the analysis technique developed here 
would be useful for the same reason in other language 
processing applications, for example specialized inter- 
active applications. 
One direction in which it is hoped to extend the work 
reported in this paper is in enhancing the capabilities of 
natural language processing systems for coping with 
idioms. Intuitively, some sort of pattern matching 
seems to be appropriate for analysing idioms (see e.g. 
Wilensky and Arens, 1980). In the context of a parsing 
system using an analysis hierarchy the patterns for 
idioms would be placed as most specialized patterns 
(i.e. leaves). LDOCE entries contain a wealth of infor- 
mation on idiomatic uses of words, and the meanings of 
idioms are expressed using the restricted definition 
vocabulary. It is hoped to extend the definition analysis 
system so that it would attempt to generate appropriate 
phrasal patterns when it encountered a definition for an 
idiomatic use of a word. It may then be possible to use 
the generated pattern and the definition of the idiom to 
produce a paraphrase of an input sentence before fur- 
ther processing takes place. In any case, a comprehen- 
sive treatment of dictionary entry analysis for language 
understanding systems clearly needs to take account of 
idiomatic word usage. 
ACKNOWLEDGMENTS 
I am very grateful to the Longman Group for making 
this work possible by granting me access to the LDOCE 
tape for research purposes. I would also like to thank 
Bran Boguraev, Ted Briscoe, and two anonymous Com- 
putational Linguistics referees for many helpful com- 
ments that improved earlier drafts of this paper. 

REFERENCES 
Alshawi, H. 1987 Memol~y and Context Mechanisms for Language 
Interpretation. Cambridge University Press, Cambridge, England. 
Alshawi, H.; Boguraev, B.; and Briscoe, E. 1985 Towards a Dictio- 
nary Support Environment for Real Time Parsing. In Proceedings 
of the Second Conference of the European Chapter of the Asso- 
ciation for Computational Linguistics, Geneva: 171-178. 
Amsler, R. 1981 A Taxonomy for English Nouns and Verbs. In 
Proceedings of the 19th Annual Meeting of the Association for 
Computational Linguistics, Stanford, California: t33-138. 
Bobrow, R.J.; and Webber, B.L. 1980 Knowledge Representation for 
Syntactic/Semantic Processing. In Proceedings of the First AAAI 
Conference, Stanford, California: 316-323. 
Calzolari, N. 1984 Detecting Patterns in A Lexical Database. In 
Proceedings of the lOth International Conference on Computa- 
tional Linguistics, Stanford, California: 170-173. 
Carnap, R. 1952 Meaning Postulates. Philosophical Studies 3: 65-73. 
DeJong, G. 1979 Prediction and Substantiation: A New Approach to 
Natural Language Processing. Cognitive Science 3: 251-273. 
Hayes, P.J.; and Mouradian, G.V. 1981 Flexible Parsing. American 
Journal of Computational Linguistics 7(4): 232-242. 
Katz, J.J.; and Fodor J.A. 1963 The Structure of a Semantic Theory. 
Language 39: 170-210. 
Kwasny, S.C.; and Sondheimer, N.K. 1981 Relaxation Techniques 
for Parsing Grammatically Ill-Formed Input in Natural Language 
Understanding Systems. American Journal of Computatiohal Lin- 
guistics 7(2): 99-108. 
Mark, W. 1981 Representation and Inference in the Consul System. In 
Proceedings of the Seventh International Joint Conference on 
Artificial Intelligence, Vancouver, British Columbia: 375-381. 
Michiels, A. 1982 Exploiting a Large Dictionary Database. Ph.D. 
thesis, Universit6 de Liege, Liege. 
Parkinson, R.C.; Colby, K. M.; and Faught, W. S. 1977 Conversa- 
tional Language Comprehension Using Integrated Pattern- 
Matching and Parsing. Artificial Intelligence 9:111-134. 
Procter, P., Ed., 1978 Longman Dictionary of Contemporary English. 
Longman Group Limited, Harlow and London. 
Pulman, S.G. 1984 Limited Domain Systems for Language Teaching. 
In Proceedings of the lOth International Conference on Compu- 
tational Linguistics, Stanford, California: 84-87. 
Schmolze, J.G.; and Lipkis, T.A. 1983 Classification in the KL-ONE 
Knowledge Representation System. In Proceedings of Eighth 
International Joint Conference on Artificial Intelligence, Karls- 
ruhe: 330-332. 
Walker, D.; and Amsler, R. 1986 The Use of Machine Readable 
Dictionaries in Sublanguage Analysis. In Analyzing Language in 
Restricted Domains, edited by Ralph Grishman and Richard 
Kittredge. Lawrence Erlbaum Associates, Hillsdale, New Jersey: 
69-83. 
Weischedel, R.M.; and Black, J.E. 1980 Responding Intelligently to 
Unparsable Inputs. American Journal of Computational Linguis- 
tics 6(2): 97-109. 
Wilensky, R.; and Arens, Y. 1980 PHRAN -- A Phrasal Natural 
Language Understander. In Proceedings of the 18th Annual Meet- 
ing of the Association for Computational Linguistics, Philadel- 
phia, Pennsylvania: 117-121. 
Wilks, Y. 1975 An Intelligent Analyser and Understander of English. 
Communications of the ACM 18: 264-274. 
