FR EE ADJUNCTS 
NATURAL LANGUAGE INSTRUCTIONS* 
Bonnie Lynn Webber Barbara Di Eugenio 
Department of Computer and Information Science 
University of Pennsylvania 
Philadelphia PA 19104-6389 
ABSTll.ACT 
In thi,~ paper, we give a brief account of our 
project Animation from Instructions, the view of 
instructions it reflects, and the semantics of one 
construction - the free adjunct - that is common 
in Natural Language instructions. 
Introduction 
Over the past few years, we have been developing a 
system for creating animated simulations from Natural 
Language instructions. When the system is complete, 
E;hese animated simulations will combine: 
® animated agents which demonstrate the instruc- 
tions being carried out in a specified environment; 
o Natural Language narration which explains what 
is being done and why. 
Such narrated simulations can then be used in instruct° 
ing agents ot' varying capacities in how to perform tasks 
with varying demands in workplaces of varying layout. 
In \[2\], we argue that the only way to create such 
blarrated simulations is to drive both animation and 
bmrration fl'om a common representation that embod- 
ies the same conceptualization of tasks and actions as 
Natural Language itself. 1 Wc also argue the difficulty 
of hand-tooling such a representation for each task to 
be demonstrated and explained. Instead, we argue for 
enabling a system to create these representations for 
itself, from Natural Language Instructions. In fact, 
we make the stronger claim that creating task anima- 
tion from anything but direct graphical manipulation 
jorces one to Natural Language as the only instruction 
~ource accessible to users other than manually skilled 
(or programming-wise) animators. 
Creating task animations from Natural Language in- 
~tructions forces us to understand instructions compu- 
rationally. Instructions as a type of text have not been 
*We thank Mark Steedman, Hans Karlgren and Breck Bald- 
win for comments and advice. They are not to blame for any er- 
~-ors in the translation of their advice into the present form. The 
,:esem'ch was supported by DARPA grant no. N0014-85-K0018, 
and ARO grant no. DAAL03-89-C0031. 
1Tiffs is not to suggest that animation can be driven solely 
from that common representation: other types of knowledge axe 
clearly needed as well - including knowledge of motor skills and 
other performance characteristics. 
studied as much as narratives as a way of describing 
tasks, but it is clear that they differ: when a na'~'ca - 
tire describes a task, it tells what happened when the 
task was performed in a particular circumstance. In- 
structions, on the other hand, commonly specify how to 
perform the task in a wide range of circumstances that 
may change during the course of performance in quite 
different ways. This has at least two consequences: (1) 
to understand instructions, one has to understand how 
instructions relate to intended behavior, and (2) in pro- 
cessing instructions, one has to deal with constructions 
that either only rarely appear in narrative or play dif- 
ferent roles than they do in narrative. 
In this paper, we start by presenting what we take 
to be the relationship between instructions and behav- 
ior, and then explore one construction often found in 
instructions - free adjuncts - explaining them in light 
of this relationship. 
2 Instructions 
Our view of instructions derives from a view of plans 
variously advocated in Pollack \[7, 8\], Suchman \[11\], and 
Agre and Chapman \[1\]. 
Pollack contrasts two views of plan: plan as data 
structure and plan as mental phenomenon. (The for- 
mer appears to be the same view of plans that Agre and 
Chapman have called plan as program.) Plans produced 
by Sacerdoti's NOAH system \[9\] are a clear example of 
this plan as data structure view. Given a goal to achieve 
(i.e., a partial state description), NOAH uses its knowl- 
edge of actions to create a data structure (a directed 
acyclic graph) whose nodes represent goals or actions 
and whose arcs represent temporal ordering, elabora- 
tion, or entailment relations between nodes. This data 
structure represents NOAH's plan to achieve the given 
goal. 
As Suchman points out \[11\], NOAH's original intent 
was to provide support for novice human agents in car- 
rying out their tasks. Given a goal that an apprentice 
was tasked with achieving, NOAH was meant to form 
a plan and then use it to direct the apprentice in what 
to do next. To do this, it was meant to generate a 
Natural Language instruction corresponding to the ac- 
tion associated with the "current" node of the graph. If 
the apprentice indicated that he didn't understand the 
1 395 
instruction or couldn't perform the prescribed action, 
NOAH was meant to "move down" the graph to direct 
the apprentice through the more basic actions whose 
performance would entail that of the original. The re- 
sult is a sequence of instructions that corresponds di- 
rectly to the sequence of nodes encountered on a par- 
ticular graph traversal. 
Pollack contrasts the above with a plan as mental 
phenomenon view, in which having a plan to do some 
action/? corresponds roughly to 
* a constellation of beliefs about actions and their 
relationships; 
* beliefs that their performance, possibly in some 
constrained order, both entails the performance of 
/? and plays some role in its performance; 
• an intention on the part of the agent to act in ac- 
cordance with those beliefs in order to perform/?. 
With respect to such beliefs, Pollack draws a three- 
way distinction between act-types, actions (or acts) and 
occurrences. Act-types are, intuitively, types of actions 
like playing a chord, playing a D-major chord, playing 
a chord on a guitar, etc. Act-types, as these exam- 
ples show, can be more or less abstract. Actions can 
be thought of as triples of act-types, agents, and times 
(relative or absolute intervals) like Mark playing a D- 
major chord last Sunday afternoon on his Epiphone. 
Because it is useful to distinguish an action from its 
occurrence in order to talk about intentions to act that 
may never be realized, Pollack introduces a separate 
ontological type occurrence that corresponds to the re- 
alization of an action. (Pollack represents an occurrence 
as OCCUR(/?), where/? is an action. Thus an occur- 
fence inherits its time from the associated time of its 
argument.) 
Agents can hold beliefs about entities of any of these 
three types: 
• act-types - An agent may believe that playing a D- 
major chord involves playing three notes (D,F~ and 
A) simultaneously, or that s/he does not know how 
to perform the act-type playing a D-major chord 
on a guitar, etc. Any or all of these beliefs can, of 
course, be wrong. 
• actions - An agent may believe that some action 
oe 1 must be performed before some other action 
a2 in order to do action /71 or that a2 must be 
performed before c~1 in order to do/?~. Here too, 
the agent's beliefs can be wrong. (It was to allow 
for such errors in beliefs and the Natural Language 
questions they could lead to that led Pollack to this 
Plan as Mental Phenomenon approach.) 
• occurrences- An agent may believe that what put 
the cat to sleep last Sunday afternoon was an over- 
dose of catnip. S/he may also have misconceptions 
about what has happened. 
Therefore one can take the view that instructions are 
given to an agent in order that s/he develops appro- 
priate beliefs, which s/he may then draw upon in at- 
tempting to "do /?". Depending on the evolving cir- 
cumstances, different beliefs may become salient. This 
appears to be involved in what Agre and Chapman \[1\] 
and what Suchman \[11\] mean by using plans as a re- 
source. Beliefs are a resource an agent can draw upon 
in deciding what to do next. 
Given this view of plan as mental phenomenon, we 
can now consider possible relationships between in- 
structions and behavior. At one extreme is a direct re- 
lationship, as in the game "Simon Says", where each 
command ("Simon says put your hands on your ears") 
is meant to evoke particular behavior on the part of the 
player. That is, 
Instruction =# Behavior 
The fact that such instructions are given in Natural 
Language is almost irrelevant. We have already demon- 
strated \[4\] that they can be used to drive animated 
simulations. Key frames from such a demonstration of 
two agents (John and Jane) at a control panel following 
instructions that begin 
John, look at switch twf-1. 
John, turn twf-1 to state 4. 
Jane, look at twf-3. 
Jane, look at tglJ-1. 
Jane, turn tglJ-1 on. 
are shown in Figure 1. 
In contrast, instructions can depart from this simple 
direct relation in many ways: 
1. Multiple clauses may be involved in specifying the 
scope or manner of an intended action. For example, 
the intended culmination of an action may not be what 
is intrinsic to that action, but rather what is taken to 
be the start of the action prescribed next. 2 Consider 
the following instructions that Agre \[1\] gave to several 
friends for getting to the Washington Street Subway 
Station. 
Left out the door, down to the end of the 
street, cross straight over Essex then left up 
the hill, take the first right and it'll be on your 
left. 
While the action description "\[go\] left up the hill" has 
an intrinsic culmination (i.e., when the agent gets to 
the top of the hill), it is not the intended termination 
of the action in the context of these instructions. Its 
intended termination is the point at which the action of 
"taking the first right" commences - that is, when the 
agent recognizes that s/he has reached the first right. 
In Section 3, we will provide many more examples of 
this feature of instructions. 
2. Instructions may describe a range of behavior ap- 
propriate under different circumstances. The agent is 
2This is not the case in "Simon Says" type instructions, where 
each action description contains an intrinsic culmination \[6\]. 
396 2 
Figure h Control Panel Animation 
o,dy meant to do that which s/he recognizes the situa- 
tion as demanding during its performance. For exam- 
ple, the following are part of instructions for installing 
a diverter spout: 
Diverter spout is provided with insert for 1/2" 
pipe threads. If supply pipe is larger (3/4"), 
unscrew insert and use spout without it. 
Here, the relevant situational features can be deter- 
mined prior to installing the spout. In other cases, they 
may only be evident during performance. For example, 
the following are part of instructions for filling holes in 
plaster over wood lath: 
If a third coat is necessary, use prepared joint 
compound from a hardware store. 
Here, the agent will not know if a third coat is nec- 
essary until s/he sees whether the first two coats have 
produced a smooth level surface. 
3. As in the plan as data structure model, instruc- 
tions may delineate actions at several levels of detail or 
in several ways. For example, the following are part of 
instructions for filling holes in plaster where the lath 
has disintegrated as well as the plaster: 
Clear away loose plaster. Make a new lath 
backing with metal lath, hardware cloth, or, 
for small holes, screen. Cut the mesh in a rect- 
angle or square larger than the hole. Thread a 
4- to 5- inch length of heavy twine through the 
center of the mesh. Knot the ends together. 
Slip the new lath patch into the hole ... 
Here the second utterance prescribes an action at a 
gross level, with subsequent utterances specifying it in 
more detail. 
4. Instructions may only provide circumstantial con- 
straints on behavior but not specify when those circum- 
stances will arise. For example, the following comes 
from instructions for installing wood paneling: 
When you have to cut a sheet \[of paneling\], 
try to produce as smooth an edge as possi- 
ble. If you're using a handsaw, saw from the 
face side; if you're using a power saw, saw 
from the back side. Otherwise you'll produce 
ragged edges on the face because a handsaw 
cuts down and a power saw cuts up. 
Such cases as these illustrate an indirect relation be- 
tween instructions and behavior through the interme- 
diary of an agent's beliefs and evolving plan. That is, 
Instructions ==~ Beliefs ¢=~ Plan ¢=~ Behavior 
3 Free Adjuncts 
In the previous section, we noted that multiple clauses 
may be involved in specifying an intended action, us- 
ing this as evidence for our view of an indirect rela- 
tionship between instructions and behavior. Here, we 
discuss one multiple-clause construct in more detail - 
the free adjunct - since it also provides evidence for 
our claim that the representation driving narrated ani- 
mations should embody the same conceptualization of 
tasks, actions and events as Natural Language itself. 
A free adjunct is defined as a nonfinile predicative 
phrase with the function of an adverbial subordinate 
clause \[10\]. It may be headed by a noun, adjective, 
prepositional phrase, or verb 3. Here we focus on free 
adjuncts headed by progressive gerundives, as they are 
quite common in instructions - e.g., the underlined 
clause in Ex. 1: 
Ex. 1 Pour mixture over cheese in casserole, 
.slgreading evenly. 
Stump notes of free adjuncts that their logical connec- 
tion with the clause they modify is not overtly specified 
\[10\] 4. Here we argue that (1) instructions exploit three 
3Constructions headed by subordinating conjunctions and 
containing a nonfmite verb, such as while fightin9 in France, he 
wan $aken prisoner are not considered to be free adjuncts by 
Stump \[10\], who calls them augmented adjuncts. 
4Free adjuncts are just one kind of a larger class of syntactic 
forms, absolute constructions, that have this property: for a more 
thorough discussion, see \[10\]. 
3 397 
logical connections between a gerundive adjunct and 
its matrix clause; and (2) to represent these relations 
requires a representation with a temporal ontology at 
least as rich as that proposed in \[6\], as well as support 
for generation relations \[5\] (defined below) and abstrac- 
tion. We conclude by showing that the role adjuncts 
play in instructions differs from the role they play in 
narratives. 
3.1 Data Analysis 
We collected 97 consecutive instances of gerundive ad- 
juncts (here called simply "adjuncts") in instructionsfl 
The syntactic structure of sentences containing these 
adjuncts is generally limited to a main clause, preceded 
and/or followed by an adjunct. The main clause de- 
scribes an action, which we call amain; #a~ will refer 
to the semantic content of the adjunct. We found that 
our corpus divided into three classes, depending on the 
logical connection between the adjunct and amain: 
1. it may augment the description of amain; 
2. it may describe a second action aa~0 that generates 
or is generated by amain; 
3. it may describe an independent action aa~ that 
should be performed simultaneously with areal,. 
It is important to remember, in the following discussion, 
that (following Pollack \[7, 8\]) an action, like an act-type, 
is a descripiion, not something in the world. 
3.1.~ Augmentation 
About half the adjuncts in our corpus supply features 
of amain, such as its starting point; necessary tool(s) or 
material(s); objects tha~ amain may create, etc. Thus, 
Cemain is a less specific version (i.e., an abstraction) 
of the intended action c~ that results from combining 
C~main and #and. For example, in Ex 2, the adjunct 
specifies the tool to use: 
Ex. 2 Using a coping or back saw, carefully cut all 
pieces to the dimensions given in the materials list. 
Alternatively, the adjunct can provide features of the 
world that have to either remain or become true after 
executing amain. 
Ex. 3 Sew the head front to back, leaving 
the neck edge open. 
The adjunct can alternatively specify a constraint on 
the execution of amain , including: 
• a manner constraint, that amain be executed in 
such a way that a state is brought about which 
continues to hold during its execution. In the fol- 
lowing example, while the agent executes the cut- 
ting action, s/he has to stay to the outside of the 
line: 
5Data were collected from five magazines - two of which de- 
scribe wood projects, and the other three, "crafts" - and one 
chapter of a "how to" book on installing wall coverings. 
Ex. 4 Transfer pattern to top back board A and 
using a jig or a scroll saw carefully cut out pattern 
staying to the outside of the line. 
a side-effect constraint, that a possible side effect 
of amain should be avoided. Verbs like take care, 
be careful, make sure etc. followed by not to ..., 
are often used: 
Ex. 5 Cut under eaves of cabin 'with chisel, 
b.ein~ care\[ul not to chip ,,oo~ 
The need to represent the result of augmentation and 
the relation between amain and a is one reason for re- 
quiring our system to have a representational capacity 
at least rich enough to represent partial descriptions of 
actions and an abstraction relation between them. 
Partial description is not meant to imply partial with 
respect to some fully specified description. On the con- 
trary, we do not assume that there is an a priori fixed 
set of features belonging to an action. To say that an 
adjunct conveys additional features of amain, does not 
mean that one can specify beforehand what all those 
features might be. 
To a first approximation, the relation between de- 
scriptions could be stated in terms of the amount of 
information that a description conveys. Note that this 
does not have to be new information: in Ex 2, the inforo- 
mation conveyed to an expert carpenter by the adjunct 
is probably redundant, given that he knows what kinds 
of saws to use. 
~.1o~ Generation 
Goldman \[5\] defines generation as that relation between 
actions intbrmally conveyed by the preposition by in 
expressions such as "agent G does fl by doing 3'" - 
e.g., "John turns on the light by flipping the switch". 
Free adjuncts can specify a generation relation between 
actions amai,~ and an4/ in either direction, without an 
overt by - for example, 
Ex. 6 As you work, clean the surface thoroughly 
each time you change grits, vacuum!n 9 off all.t.h.e 
_dust and wiping the wood with a rag dampened with 
.turpentine or paint .thinner. 
\[aa~ GEN amain\] 
Ex. 7 Cut one 7xT-inch square from foil. Fold cor- 
ners to center of square; cut in half on the diagonal 
creating two triangles. 
\[amain GEN aa~/\] 
Ex. 8 Sew bottom canvas bag to bottom of front 
and back, makin~l a long rectanfle. 
\[amain GEN aa~\] 
4 398 
In the case of generation, only one action is executed 
per se, generating the other as a result. 
One natural question to ask is why two different de- 
scriptions are given of the same action. The reasons 
are the same as in any text: to make explicit the pur- 
pose of ~Jt action or a salient feature. For example, in 
Ex. 6, clean provides a unifying description for the two 
actions expressed in the adjuncts, and by doing so, in- 
dicates their purpose. In Ex. 7, the result of amain (the 
two triangles) is mentioned explicitly, in order to intro- 
duce these new referents in the agent's discourse model. 
In Ex. 8, the description a long rectangle provides a vi- 
sual clue to the result to be achieved. (This may be an 
additional purpose for the generate relation in Ex. 7 as 
well.) 
Again, Ex. 6 shows the need for abstraction in our 
representation, in the form of one or more abstraction 
hierarchies of action descriptions: to understand this 
example, we need to know that both vacuum and wipe 
are specializations of clean. 
3.1.3 Simultaneity 
If the adjunct describes an action oza4/ that is indepen- 
dent of oqnai,, it is meant that both are to be executed 
simultaneously: 
Ex. 9 Soak paper in water for 1 hour; remove pa- 
per, then smooth onto board, squeezing out excess 
.~d ir~ble_~ Staple paper to board along the 
edges. Mix rose madder and water; pour onto wet pa- 
per, tilting board to spread color. 
:Ex. 10 Unroll each strip onto the wall, 
the foil into place vertically (not side to sidS_ to avoid 
warping and curlinq__at the edq~es. 
3.2 Aspect and Event Structure 
Earlier, we claimed that the representation driving nar- 
rated animations should en'lbody the same conceptual- 
ization of ~asks, actions and events as Natural Language 
itself. We take the conceptualization of actions and 
¢;vents to be the tripartite event structure described by 
Moens and Steedman (hereafter, M~S) in \[6\]. 
The goal in \[6\] is to provide a single explanation of 
aspectual profiles, of changes in aspectual profile re- 
lated to the use of adverbial and prepositional mod- 
ifiers, and of the purported temporal "ambiguity" of 
when-clauses. The explanation makes use of a tripartite 
event structure which M~S call a nucleus. A nucleus 
consists of a preparatory process, a culmination and a 
consequent state. Within this framework, an event de- 
.'~cription interpreted as a PROCESS corresponds simo 
ply to a preparatory process, while a CULMINATED 
PROCESS corresponds to an entire nucleus. CULMI- 
NATIONS (Vendler's achievements \[12\]) correspond to 
a culmination followed by its consequent state. 
Within this framework, M~S attribute changes in 
~!~spectual profile brought about by modifiers (viewed as 
functions from event description to event description) 
to two factors: (1) The modifier, viewed as a function, 
may have a different output type than its input type. 
The modified form will thus have the same aspectual 
type as the function's output. (2) When a function 
demands a different aspectual type for its input than it 
has been given, a mechanism called coercion maps the 
input to the needed type. This may change semantic 
features of the input, before function application. 
What we shall show here (rather briefly) is that this 
same tripartite nucleus can ground the possible inter- 
pretations of augmentation (Section 3.1.1) and simul- 
taneity (Section 3.1.3), and in fact, account for ambi- 
guities in interpretation. We start with the following 
minimal pair: 
Ex. 11 Starting with this mark, make another mark, 
leaving exactly P inches between marks. 
Ex. 12 Starting with this mark, make a series of 
marks, ~c_tly_ 2 i nche_s between marks. 
In M&S's framework, making a (single) mark (Exam- 
ple 11) could be considered a CULMINATION. The 
plural "series of marks" in Example 12 would then map 
this singular interpretation to a CULMINATED PRO- 
CESS through iterating mark-making. (Iterated mark- 
making culminates when there is no more room to make 
marks.) The augmentation in Example 11 constrains 
the distance between the single pair of marks, that in 
Example 12, the distance between each pair of marks 
produced during the iteration. 
Now consider the following example of simultaneity: 
Ex. 13 Wire vines together at one end. Twine vines 
into an 8.inch diameter wreath, fastening 
with wire to hold. 
The second sentence mentions two independent actions 
- twining the vines into a wreath (amain) and fastening 
(aa~/). In M~cS's framework, the action amain can be 
taken to be a CULMINATED PROCESS in two differ- 
ent ways: a single homogeneous twining process, which 
culminates when one has used up all the vines, or (as 
above) an iteration of individual twinings, cuhninating 
for the same reason. In the first case, fastening happens 
at the single point of culmination - its purpose being to 
prevent the two ends of the wreath from coming apart. 
In the second, fastening happens at the end of each iter- 
ation - its purpose being to keep the strands together. 
To capture both these interpretations (and decide be- 
tween them) requires a representation such as M~S's 
rich enough to capture the required event structure. 
3.3 Relation to Previous Work 
The most extensive analysis of the semantics of free 
adjuncts (in English) that we are aware of is that done 
by Greg Stump \[10\]. However, all his examples come 
from narrative text, and as a result, he focusses on their 
truth-conditional properties. For example, he draws a 
distinction between strong and weak adjuncts: 
5 399 
Ex. 14 a) Having unusually long arms, 
John can touch the ceiling. 
b) Standing on the chair, 
John can touch the ceiling. 
Ex. 15 a) Being a businessman, Bill smokes cigars. 
b) Lying on, the beach_, Bill smokes cigars. 
Stump calls the adjuncts in both a sentences strong, 
because their actual truth is uniformly entailed. He 
calls those in the b sentences weak, because their actual 
truth can fail to be entailed. 
Related to this, Stump also notes a Causal flavor in 
strong adjuncts. Consider the adjuncts in the a sen- 
tences in both Exs. 14 and 15. The sense is that in 
both cases, the main clause assertion is true because 
the adjunct is. Weak adjuncts, on the other hand, have 
a conditional sense: it is (only) when the condition de- 
scribed in the adjunct is true that the main clause as- 
sertion is true. 
While these observations appear to be both correct 
and relevant in narratives, this strong/weak distinction 
appears :irrelevant for instructions, which do not con- 
cern themselves with truth conditions in the same way 
as narratives. The only thing in instructions that comes 
close to the conditional sense of weak adjuncts is the 
perfective gerundive adjunct, as in 
Ex. 16 Having..basted the seams, check again for fit. 
Such adjuncts do convey a similar sense that it (only) 
when the action described in the adjunct is complete 
that the main clause command is relevant. 
In Section 3.1, we essentially tried to show that in 
instructions, gerundive adjuncts play a role in further 
specifying intended action. They may do this through 
augmenting amain, through providing an alternative de- 
scription of Otrnai n through generation, or through spec- 
ifying another (independent) action that must be per- 
formed simultaneously with Otmainin some way. Thus 
we conclude that gerundive adjuncts (if not all free ad- 
juncts) play a different role in instructions than they 
do in narrative text. This emphasizes the importance 
of analysing constructions in situ, rather than assum- 
ing that conclusions based on narrative text will hold 
equally of instructions. 
4 Summary 
In this paper, we have given a brief account of our 
project Animation from Instructions, the view of in- 
structions it reflects, and the semantics of one particu- 
lar construction that occurs often in the type of instruc- 
tions we will be handling. The project is proceeding on 
several fronts, including the following: (1) Similar anal- 
yses are being done of other constructions that com- 
monly occur in instructions \[3\]; (2) we are starting to 
develop a representation that embodies both the tem- 
poral ontology \[6\] that grounds the semantics of these 
constructions and an abstraction mechanism- notice 
that when we talk about abstraction we do not limit 
ourselves to abstraction hierarchies: we intend abstrac- 
tion as a general relation between more and less specific 
descriptions of actions; and (3) translation processes 
are being expanded for mapping that representation 
into forms that our simulation system \[4\] can deal with. 
More detailed description of the system as a whole is 
given in \[2\]. 

References 

\[1\] Phillip Agre and David Chapman. What are 
Plans For? A.I. Memo 1050a, Artificial Intelli- 
gence Laboratory, MIT, October 1989. 

\[2\] Norman Badler, Bonnie Webber, Jeff Esakov and 
Jugal Kalita. Animation from Instructions. Mak- 
ing Them Move: Mechanics, Control and Anima- 
tion of Articulated Figures. Morgan-Kaufmann, 
1990. 

\[3\] Barbara Di Eugenio and Bonnie Webber. Ac~ 
tion Specifications in Natural Language Instruc- 
tions. Technical Report, Dept. of Computer 
Information Science, University of Pennsylvania, 
Philadelphia PA. Forthcoming. 

\[4\] Jeffrey Esakov and Norman I. Badler. An Archi- 
tecture for Human Task Animation Control. In 
Knowledge-Based Simulation: Methodology and 
Applications P.A. Fishwick and R.S. Modjeski 
(eds.), Springer Verlag, New York, 1989. 

\[5\] Alvin Goldman. A Theory of Human Action. New 
York: Prentice-Hall, 1970. 

\[6\] Marc Moens and Mark Steedman. Temporal On- 
tology and Temporal Reference. Computational 
Linguistics. 14(2), 1988, pp. 15-28. 

\[7\] Martha Pollack. Inferring Domain Plans in 
Question-Answering. PhD Thesis, Dept. of 
Computer and Information Science, University 
of Pennsylvania, Philadelphia PA. (Available as 
Technical Report MS-CIS-86-40, University of 
Pennsylvania, May 1986.) 

\[8\] Martha Pollack. Plans as complex mental 
attitudes. In Intentions in Communication, 
J. M. P. Cohen and M. Pollack, Eds., MIT Press, 
1990. 

\[9\] Earl Sacerdoti. A Structure for Plans and Behav- 
ior Elsevier, New York, 1977. 

\[10\] Greg Stump. The Semantic Variability of Abso- 
lute Constructions. Dordrecht: D. Reidel, 1985. 

\[11\] Lucy Suchman. Plans and Situated Actions: The 
problem of human machine communication. Cam- 
bridge University Press, 1987. 

\[12\] Zeno Vendler. Linguistics and Philosophy. Ithaca 
NY: Cornell University Press, 1967. 
