The Role of Inversion and PP-Fronting 
In Relating Discourse Elements: 
some implications for cognitive and computational models of 
Natural Language Processing 
Mark Vincent LaPnlla 
The Artificial Intelligence Laboratory 
and 
The Department of Linguistics 
University of Texas at Austin, 
Austin, Texas USA 
11 April 1986 
O. Abstract 
1 This paper will explore and discuss the less obvious ways 
syntactic structure is used to convey information and how this 
information could be used by a natural language database 
system as a heuristic to organize and search a discourse space. 
The primary concern of this paper will be to present a 
general theory of processing which capitalizes on the 
information provided by such non-SVO word orders as 
inversion, (wh) clcfting and prepositional phrase (PP) fronting. 
since it seems that these non-SVO structures are sensitive to 
NPs. Thus every discourse representation is in some way 
redundantly specified for at least one constituent. 
For example, the organization of the discourse representation 
for the sentences, "In the forest stood a house" and "In the 
park, Mary kissed John", are: 
Labcl: house, forest 
D.R. : a house stood in the forest 
and 
1. Introduction 
English at its simplest is an SVO, Subject Verb Object, 
language. However it is not limited to SVO order. Clefts, 
pseudo-clefts, inversion, topicalization, left dislocation and 
various types of fronting are instances of deviation from SVO 
order. Non-SVO orders are not exceptional or found only in 
obscure literary writing. They abound throughout writing and 
speech. An interesting question is what use do these structures 
do in English, and how can a cognitive or computational 
theory use such information? 
Non-SVO word order helps the reader (or listener) to 
construct a discourse representation. It is a heuristic devise for 
creating coherent and cohesive representations of text and for 
searching existing representations. In other words, it is a 
device for finding in long or medium term memory the 
relevant context, or discourse space, in which to embed and 
interpret the sentence being processed. It is a linguistic device 
used for changing the discourse focus (Sidncr 1978, 1983; 
Grosz 1978, 1981). It is important to note that no particular 
discourse representation construction schema is assumed in 
this claim but rather that this claim holds no matter what 
type of construction algorithm is used. Also, what is being 
proposed here is not a linguistic rule for constructing discourse 
representation but rather a principle (heuristic) for 
constructing them. That is a principle for organizing and 
searching discourse representations. 
As a sentence is processed a representation is made of it. 
This representation consists of a label, which is a (syntactic 
and semantic) parse 2 of the first constituent cncountercd, 
followed by a parse of the whole sentence. Actually, one could 
have multiple labels that consisted of the NPs in the sentence, 
1This research was supported by the U.S. Army Research Office 
under contract DAAG29-84-K-0060. Artificial Intelligence 
Laboratory, Department of Computer Sciences. 
I would like to thank Professor Robert F. Simmons for his 
support, help and criticisms. 
2The exact representation of the text will be discussed later. For 
now English Will be used to represent what eventually will be a 
discourse representation of some sort. Parse is being used hcre in a 
general sense to mean "give the structure of". The "structure" of 
course depends on the linguistic theory used to give the parses. 
Label: Mary, Johu, park 
D.R. : Mary kissed John in the park 
Two uses could be made of this system of "labels"• The first 
is to simply use the labels to index the text and to facilitate 
the search through the text. The second would be to use the 
labels as nodes in a semantic network. Thus in the discourse 
fragment, "A house stood in the forest. Outside stood an 
angel." the label "outside "3 could be related to either the 
label "house" or "forest" or both via a "location" arc. 
Due to tile lack of space the former (index) nsc will be tile 
focus of this paper but note that it is indeed difficult to 
separate these two uses since it is necessary to locate ttle 
material necessary to construct a discourse space, i.e. even if 
one were only using a very restricted procedure for creating 
discourse representations one would still need to locate 
previous discourse items in order to resolve anaphora, and 
therefore create some sort of structured link from discourse to 
discourse. 
In the sections that follow we will give linguistic motivation 
for this analysis. Due to the enormity of the urea and the 
limitation of space, only inversion and PP-fronting will be 
discussed. In the final section we will present the findings of a 
study done to see if this system of "labels" could be used to 
speed the construction of discourse representations, specifically 
the resolution of anaphora. 
2. Inversion 
In Green's (1980) study of the discourse function of various 
"classes" of inversion, she assigns each instance of inversion a 
particular function. This approach, however, is bound to bc 
inadequate. If one tries to compile a list of such functions for 
various syntactic fornls in a language, how does one know if 
the list is complete? Every time a new function is discovered 
for a form one must add it to the list. Another problem with 
this approach is that one never knows if one has been specific 
enough or gcncral enough. Being too specific in the 
characterization of such functions creates a very long list and 
3Or rather the relation "outside" that implies outside of 
something. 
168 
can miss generalizations. Being too general might hide tile 
correct function of a form in a language. This section will 
review and criticize a subset of her list of proposed hmctions. 
2.1. Use 1 
inversion allows the listener first to identify tile object being 
talked about before assigning information to it, whether new 
or old. This is most noticeable in the speech of sportscasters 
(Green; p 584): 
a. Underneath is Smith 
b. IIigh in the, air to get the ball was Jim Brady. 
c. Bringing the ball up is Marty Mcstemacher. 
d. Back come the Kahoks with the ball. 
c. And sitting down is Kcvin ,loses. 
In this case inversion allows the viewer to single out the player 
on the TV screen before processing his name. This allows the 
viewer to first pick out the player, makca non-linguistic 
representation in memory of that player and then add the 
(new) information given by the sportscaster: the player's 
name. If inw;rsion were not used the viewer would \[lave to 
first store the namc given by the sportscaster, idcntify the 
player, construct a representation of that play and his actions, 
retrieve the player's name and then assign that information to 
tile representation created by the description of tile player's 
actions. This is a much more difficult and time consuming 
procedure. In this itmtancc, inversion hclI)s to cut down the 
amount of processing necessary to construct a representation. 
Therefore one eoukl hypothesize that upon hearing/reading 
the first few elements of a sentence, the listener follows three 
steps: 
1. if the sentence is SVO (and does not have any overt 
signals to search for a previously mentioned item of 
information,) construct a representation of tile sentence 
and add it to the local discourse space. 
2. Else search memory for the last mention of the 
item under construction and add the "new" information, 
i.e. what is in the predicate, to that local discourse 
space. Pointers are left pointing to both discourse 
spaces. 
3. Otherwise, construct an entirely new discourse 
space. 
Actually it is nnclear where the information should be 
deposited. For example, a house has been robbed. The police 
investigate. They ~ksk questions concerning tile robbery. Then 
the father of the household, when asked who they think could 
have done it, says: "As I said before, that boy John is a 
thief." 
"\['he police who do not have any previously knowledge of 
John add this information to tile present discourse space. They 
probably note that the person who said it has reasons for 
bringing it to tile listener's attention that he had said this 
before, llis daughter, however, who knows John but did not 
know her father's opinion of him, adds to her mental 
representation labeled "John" this information. She also adds 
this utterance to the cun'ent discourse space, i,e. "robbery of 
house". The man's wife on the other hand already knew about 
her husband's opinion. It would seem redundant for her to add 
this information to her mental representation labeled "John". 
What she probably does is call up tile mental reprcsentation 
"John" and leave u pointer pointing to it from the current 
discoursc space "robbery of house", thereby connecting the 
two representations. So it would seem that things arc not as 
cut and dried as one would suppose. 
2.2. Use 2 
Tile second use for inversion which Green cites is its ability 
to conncct pieces of discourse together. This is used frequently 
in journalism. One can link and expand a 1)rcviously 
mentioned proposition by anaphorically referencing it ill the 
gramlnatical subject slot, thereby smoothly linking the new 
information in the predicate to thc old, prcviously mentioned 
information. She also notes that this connective function is 
uscd in literary texts or expository prose. She finds that 
inversion provides a concise form in which news writers can 
begin a sentencc with old information. 
Ill the analysis being developed ill this section, tile only 
difference between tile sportscasters' speech and this 
"connective" function is that the "connecting" in the 
sportscasters' speech is done to an image rather than a 
(previously mentioned) linguistic concept. Ill tim sportscaster 
case, the viewer ha.s to create a nonlinguistie representation in 
his consciousness, i.e. his discourse space. Then when tile name 
of the player is mentioned the viewer assigns the information 
to the representation. The "connective function" clme differs 
ouly in the type of reprcsent~ation built. 
2.3. Use 3 
in her third function, Green expounds upon the notion that 
what is invcrted is not ncccssarily new information. She shows 
that inversion can be used to set a scene for an event or as a 
means to locate actors in a story, e.g. "Outside stood an 
Angel", "Ill a little white house lived two rabbits." 
Notice how similar this "function" is to the sportscasters' 
speech (and the newspaper examples). Tile sportscaster sl)ecch 
uses inversion to identify tile player so that tile listener can 
more easily identify the (new) information, i.e. his name, with 
the player. The scene-setting, and literary connecting, function 
of inversion identifies a locale in which to place the actors, the 
characters. From a processing point of view these are the same 
things. Even Green notes tile similarities between the 
jonl'nMist hmction and the literary connecting function. 
13(b) Sprawled in the foreground is George 
169 
Price. 4 
"Ex. 13b, which is part of the description of an 
accompanying picture, is very much like the news 
examples: it identifies a piece of the ~ an 
individual.L by ~ him with reference to --mor__~e 
specifically,, in the fore ro~ of--something taken 
to be alre%.~_ ~ th__.~e ~ieture as a whole." (p. 
588;underlining added) 
In all these cases a discourse representation can be created 
that uses as its label the first constituent of the inversion. In 
the sportscasters' speech the representation of the sentence(s) 
would be linked to the image of the player and the 
(information) "living of the two rabbits" would be assigned to 
the discourse representation labeled "white house". This last 
assignment might seem strange but suppose that the house 
was previously identified as being in a large forest: 
Discourse i. 
In a large forest stood a house. 
In the house lived two white rabbits. 
This use of inversion does not seem to signal an extensive 
search of memory but rather seems to create a more local 
chain of association: there is a sense in which the second 
sentence is an elaboration of the stored information about "a 
house". Compare the short discourse structure above with 
Discourse 2. 
In a large forest stood ~ house. 
Two white rabbits lived in the house. 
Intuitively, this discourse seems harder to process than the 
previous one. Finally compare these sentences with: 
Discourse 3. 
A house stood in a large forest. 
In the house lived two white rabbits. 
This last discourse seems ~ easy to process as Discourse 1 
(D1). The theory under development here accounts for this. 
Discourse 2 (D2) is harder to process than D1 because when 
processing D2 one must store the concept "two white rabbits 
lived..", in some manner, and then search for a previous 
mention of "house" in which to embed the information. 
(Notice it is not the inversion that makes "the house" 
anaphoric but rather the use of the definite article. 5) 
4Green's numbering. I will continue to use Green's numbering for 
her examples. I will use a more coherent numbering system for my 
examples. 
5I would still like to maintain that inversion is used as a signal of 
anaphora. 
Inversion, in this case, makes the discourse easier to process. 
Discourse 3 is as easy to process as Discourse 1 because the 
inverted element, "a house", becomes the label and the 
"connecting phrase", the label, of the representation of the 
second sentence is "in the house". An interesting observation 
is that Discourse 5 seems harder to process than D1, as 
expected, but easier than D2. The explanation for the latter 
observation is not at first obvious. 
Discourse 5. 
A house stood in a large forest. 
Two white rabbits lived in the house. 
In the case of the above discourse (DS), "a house" is the 
label of the last representation built. So even though the 
connecting phrase "in the house" is not in initial position, 
which accounts for why this discourse is harder to process 
than D1, there is a '"top level" item, i.e. label, "a house", to 
which the connecting phrase "in the house" can connect. 
However, in D2 not only is the connecting phrase "in the 
house" buried in the second sentence, i.e. not in initial 
position, the item to which it must connect is also buried. 
Extending this reasoning the .theory would predict that D3 
would be easiest to process, D1 and D4 the next hardest and 
D2 the hardest. This claim is a strong claim about the internal 
structure of discourse representations and could be falsified 
with psychological experimentation. 
The intuitive sense in which a sentence is harder or easier to 
process is perhaps also related to the idea that the subject is 
an external argument which, participates in a predicate 
relationship with the entire VP and not just the verb. In this 
sense the object(s) of the verb are more "deeply embedded" in 
the sentential predicate than the subject. THus inversion, PP- 
fronting, etc., can be viewed as moving an embedded, or 
internal argument, to a more external position, e.g. adjunct 
position. 
Not also that those verbs which appear in inverted sentences 
seem to be ergative verbs. That is the deep structure of the 
sentence "Outside the house stood an angle" is probably \[S e 
\[VP stood an angle\] \[pp outside the house\] 6 (Helm 1085, 
personal communication). This might help to explain the 
greater "availability" of fronted material. 
An important point to note is that Green does not consider 
the scene setting function and the literary connecting function 
to be the same thing. In the scene setting examples the 
inverted element is completely new information, where~s in 
the literary connecting function this does not have to be the 
case. This is an important point for the theory in this paper as 
well. Crucially, the claim of this theory is not that the 
inverted element is old information but that it is the 
170 
important element with respect to embe~tding of information. 
When it is new information it sets up a context in which new 
information can be embedded, including the information in its 
own predicate. When it is old information it serves to find the 
correct context in which to embed the information in the 
predicate. 
In all of the above cases, inversion is used to locate and 
identify an (old) entity, an event in the sportscaster speech, a 
(previous) location, or all image, and give more (new) 
information about that entity, or create a context in which to 
embed information. 
3. PP Fronting 
PP-fronting is used to provide a continuity, a cohesion, in 
the text. It provides a useful progression of labels to which to 
attach the accompanying information. For example, an article 
by Lawrence (1985) opens with a fronted prepositional phrase 
which provides a tinm setting, or relation on the (narrative) 
time line, for the activity in the sentences which follows7: 
\])iscourse 6. 
befol'c I was tall enou~ to ride on the 
coaster ~, I spent many pleasant hours per- 
suading my reluctant father to accompany me. (p. 4) 
The PP also provides a way to link up the topic (theme) of tile 
article to the opening statement of the article. The tlmmc is 
the "new adult" Amusement Parks. The article initial 
prepositional phrase picks out a particular item within an 
amusement park and associates the remembrances of the 
writer to it. 
The next sentence also has a fronted PP. This PP also links 
the next sentence to the article's main topic: 
Discourse 7. 
As an aficionado of amusement ~12~., 1 was over- 
joyed when our whole family finally flew to 
California to tackle Walt Disncy's extravaganza. (p. 
4) 
The next paragraph starts out witb yet another fronted 
temporal prepositional phrase, moving the time setting up to 
the present: 
Discourse 8. 
More than two decades later, I'm still journeying 
to parks. 
These first few examples of word order deviation all have the 
characteristic of giving the reading a temporal "focus" and 
order in a series of events that occur over a number of years. 
In order for a discourse processor to understaml this text, it 
6This is roughly the structure. 
7The nnderlining is mine 
would have to have a place to start. The logical starting place 
would be with the label AMUSEMENT PARK, since this is in 
the title of the piece. Under this discourse representation label 
it could build other representations. The first representation 
that it would build would be about the author since this is the 
first matrix NP of the first sentence, D6. For the next 
sentence, D7, it would already know what type of temporal 
relation to assign to the proposition expressed in the matrix 
clause. The sentence in the next paragraph is easily processed 
since it advances the time of the preceding paragraph. Rather 
than building a representation of items and attaching to this a 
set of properties, these fronted PPs build an abstract 
representation of temporal items related by ttle time of each 
item. 
The general theme of the article is amusement parks. 
Ilowevcr, since the opening of the article is more a personal 
recollection rather than expository, tile information that needs 
to be organized is not information about particular objects, 
i.e. amusement parks, but rather episodes in the author's life. 
Each cpisode's temporal relation is specified by the fronted 
prepositional phrase. In general this is the function of fronted 
prepositional phrases, the specification of relations. This is a 
widely used tcchnique; used more than any other non-SVO 
pattern: 
The general hypothesis is that the first thing that one 
encounters in a sentcncc is a link to preceding information, 
either explicit or implied. The link provides the proper context 
in which to build tile new representation. It also provides the 
means for quickly searching the discourse space. 
PP-fronting, like inversion, allows the reader to connect the 
current phrase, or sentence, being processed to an appropriate, 
and most likely salient, antecedent. For example Green (1980) 
comes to the same conclusion about inversions in 
sportscasters' speech. She notes that sportscasters use 
inversion when broadcasting play-by-play to identify tile 
player by his action and then name him: 
~ ~ .and then ~ it was Dave Bonko. 
Back come the I(ahoks with the ball. 
An(.._~l in comes nnmber 5.~1, and that will be Mike 
Matakitis. 
Into tile ~ for the New Trier West is Brenner. 
The reason she gives is that this is helpful to tile TV viewers, 
since they don't have scorecards identifying the players. Shc 
goes on to say that in this way the viewer can single out the 
player on the screen before receiving his name. (This also gives 
the sportscaster time to look up the player's name if nccd be.) 
Further more she notes that sportscasters use this inverted 
style even when tile player arc well known or there numbers 
arc clearly visible. This observation fits in nicely with the 
modal being built here. 
171 
Thus like inversion PP-fronting is used to help link, via 
labels, i.e. focused material, one discourse representation to 
another. 
4. The Experiment 
In the above two sections we briefly motiw~ted and 
developed an analysis of the organization of discourse 
representations. Basically the analysis claimed that each 
discourse representation, no matter how it is represented, i.e. 
what particular theory or formalism, were indexed via their 
focused NPs. The analysis also claimed that non-SVO word 
structure was a signal to search through the labels to locate 
the structure in which to embed the representation currently 
being processed. 
There are two aspects of this analysis that we will focus on 
in this section: the creating of labels and the searching of the 
labels. The more complicated aspects of building and 
embedding, or relating, the strnctnrcs to one another will be 
iguored for the sake of exposition. 
A simple experiment was performed to explore the 
computational usefulness of the proposed labeling system. 
Three programs werc written in Symbolics \]?rolog. Each 
program processed a set of twenty-six sentences and created 
discourse representations. To create the discourse 
representations the DRS construction algorithm found iu 
Kamp (1986) was used. Added to this were straightforward 
rules for creating DRSs for locative prepositional phrases. The 
task for each program was to resolve simple anaphora by 
searching through the discourse representations for the 
antecedent. A straightforward feature matching technique was 
used to do this. If one were trying to resolve the reference for 
a pronoun and a full NP then only the features of the lexieal 
item, e.g. masculine, singular, was matched. If the reference 
for a full NP was being resolved then the whole lexical item 
was search for. 
The first program only constructed discourse representations. 
it did not construct labels as well. Thus whenever anaphoric 
resolution was called for by thc DRS Construction algorithm, 
this program had to search through thc cntire data basc until 
a match was found. The second program created labels but 
they were only searched when the sentences being processed 
had non-SVO structure. The third program created labels as 
well but it only searched the labels. That is the heuristic 
always applied. 
Each DRS was a flat list. Each label list was also flat. Before 
each run of the program the machine was cokl booted. 
The data was a list of 24 sentences. The l~st sentence 
contained the only fronted PP, which referred back to the first 
sentence. The results of this experiment are discussed in the 
next section. 
172 
5. Results 
In pilot experiments tile DRS list was allowcd to have 
complex structure, In other words, the DRS list was a lists of 
lists (of lists, ete). The label list on the other baad was flat, 
i.e. a list of lists. The processing in the case of the complex 
structure was speeded up by a factor of 3, overall, when the 
heuristic was applied at all times (in the third program). In 
the case of the second program only the processing of the non- 
SVO sentences, the last sentence in the corpus were sped up. 
However in subsequent experiments the DRS lists were all 
converted into flat lists. The effect was less dramatic but 
significant none the less. Below is given two tables of the 
figures for all 3 runs. The first table is the time each program 
took for processing the non-SVO sentence, i.e. the last 
sentence in the corpus and the time it took to process the 
corpus overall 8, 
I Time (secs) I 'rime for 
Table i J of Non-SV0 \[ Entire 
I Sentence I Corpus 
..................................................... 
First Program I \[ 
No heuristics \] .023 \] .66 
I I 
................................................... 
Second Program I I 
Heuristic on \[ .0187 \[ .659 
Non-SVO \[ I 
structure I I 
.............................................. 
Third Program \[ \[ 
Heuristic at I .019 \] .54 
All times. J I 
8The times given iu table 1 are rounded to tile nearest thousandth 
of a second. Ilowevcr all percentages were computed with original 
microsecond numbers and then rounded. 
I g of ~peed 
'r~b\].e 2 I Up as compared 
I to F±r~t Progra.ln 
First Program I Speed up: 
No lleurist;J.cs I O~ on NolI-SVO 
I Og overall 
Second Program I Speed up: 
Ileurl~:lt,:lc on I 19% on Non-sVO 
NolI-SVO I lg oveFa\].l 
structure \] 
Third Progra, i, \[ Speed up: 
HeurJ.~t:\].c ~t \] 16~ on No~I-.SVi} 
All timet~ \] 19% overs, l\]. 
It is obvious from tile above table thaC always searching the 
label list(s) is far superior to searching tile larger discourse 
struetm'es. I~ is also significant that whm, the heuristic 
(searching ~he label li,'.;t) is only applied l.o Non-SVO 
struetui'es~ i.e. tile fast senl, enee of the eo,'pl,s~ t;ll,'~b tilt., spee(1 
up is significaa,t. (The difference between the second a,ld third 
program wil, h respect to I, hc last sentence, the non-.SVO 
sentence is not significant and i~; due to nlaehine rel;~t.ed 
factors.) The,'efore it seems that it is worthwhile buih\[ing a 
separate list of NPs and searching that list; for at; least 
resolving a.naphora and using it not only for the (linguistically) 
motivated non.4~VO structure search but all the time azs well. 
6, Concluslon 
We have mo~ival, ed a heuristic device that consists of 
creating a llst of the NPs in each sentence. This wa~'~ 
motivated on linguistic grounds for non-SVO word or<let. It 
weal suggested that this list could facilitate in the construction 
discourse ,'epresenl, ations and for resolving auaphora in a 
computer program. The latter of these two suggestions was 
investigated. It was found that iudeed a significant deere~se 
in processing time occurred. 
The first of the two above hypothesis was ,lOt empirically 
investigated. One avenue of interesting research would be to 
see if the information provided by non-SVO word order could 
help in the construction of more complicated discourse 
rei>reseatations and if such representations would help in are~m 
like Question-Answering. 
A second avenue of research would be in psycholinguistics. 
Basically experiments could be set up to test the hypothesis 
that non-SVO word order some how signals a search of the 
discourse space. 

References 

Siduer, C. l"ocusing and Discourse. Discourse \]b'occss, Oct- 
Dec 1983, pp. 10%130. 

Chafe, Wallace. Language and Consciousness. Language, 
1974, 50, ll 1-133. 

Chafe, Wallace. Givenness, Con:restiveness, Definiteness, 
Subje.cts, Topics, and Point of View. 1,1 Charles l,i 
ted.), 5',tbject and 7bpic. New York: Academic Press, 
1 !t76. 

Davidson, Alice. Peculiar Pa~,~sives. Language, March 1980, 

Green, Ceorgia M. Some Wherefores of English Inversions. 
Language, 1980, 51,~3), 582-602. 

Grosz, B.. Focusing and Description in Natural l,anguage 
I)ialogues. In Josh:, A. B. L. Wetter and 1. A. Sag 
(F, ds.), l';lements of Di,scmtrse \[Szderstanding. 
Cambridge: (Sunbridge University Press, 1981. 

(\]rosz, tl.. Focusing in I)ialog. In Tinlap-2: Th.corclical 
l.ssues in Natural Language 15"ocessing. New York: 
ACM an<l ACI,, 1978. 

llavilangd, Susan t';. and Ilerbert Clark. WhaL's New? 
Acquiring new informal;ion im a process in 
comprehension. Journal of l/erbal Learning and l/erbal 
Behavior, 1!t74, 13, 512-.,538. 

Ih!im, h'ene R. The 5'c.rnantics of Definite and b~definite 
Noun Phrases. Doe~ol'al dissertation, University of 
Massachusel, ts at Ainherst, September 1982. 

Kemp, liens. Situations in Discourse without Time or 
Questions. CSLI, 1986. To appear in CSI,I Stalfford. 

Kemp, Hans. Context, Thought and Communication. In 7he 
Proceedings from the Aristotelian ,5'ociety. , 1985. 

Lawrence, A. Amusement Parks. AAA, 1985, . 

l'rinee, Ellen F. A Comparison of Wh-Clcfts and It-Clefts in 
Diseonrse. Language, 1978, Vol. 5"5~(4). 

Schank, Roger C. Dynamic Memory: a theory of reminding 
and learning in computers and people, Cambridge 
University Press, 1982. 

Sidner, C. l,evels of Complexil, y in Discourse for Anaphora 
l)isambiguation and Speech Act Interpretation. 1,1 
Proceeding of the Fifth, International Joint Conference 
on Artificial Intelligence. IJCAI, 1978. 
