ACTION RELATIONS 
IN RATIONALE CLAUSES AND MEANS CLAUSES* 
CECILE T. BALKANSKI 
Aiken Computation Lab, Harvard University 
Cambridge, MA 02138, U.S.A 
Abstract 
Communication among agents collaborating on a task 
often involves complex utterances about multiple ac- 
tions. In this paper, we analyze two particular multi- 
action constructions, utterances with means clauses 
and utterances with rationale clauses. We present 
the distinctive features of these constructions, de- 
scribe their logical form, and define interpretation 
rules for deriving their meaning in terms of the gener- 
ation, enablement and Contributes relations between 
actions which have been motivated independently by 
research in plan recognition. We also demonstrate that 
these rules yield the appropriate properties of ratio- 
nale clauses and means clauses and show how these 
constructions can be distinguished algorithmically. 
1 Introduction 
Communication among agents collaborating on a task 
often involves complex utterances about multiple ac- 
tions, as illustrated in the following example: 
By going and getting the key and then coin- 
ing and unlocking the grate, for that you get 
a ccrtain number of points, and then as you 
collect treasures, you get more points. \[1\] 
Multi-action utterances like these present two par- 
ticular challenges for natural language interpretation 
systems: determining an appropriate representation of 
their meaning, and defining the process by which this 
representation can be derived from natural language 
utterances. Substantial work has been done on the de- 
velopment of logical forms as a necessary step toward 
the goal of representing the meaning of an utterance. 
However, very little is yet established on the represen- 
tation of the information communicated in complex 
action utterances like the one given above, and still 
less on the interpretation of thc relations tbat connect 
the actions described in these utterances. 
In this paper, we analyze two particular multi-action 
constructions, utterances with rationale clauses and 
utterances with means clauses, illustrated respectively 
in (la) and (tb) below, with the goal of meeting, at 
*I would like to thank Barbara Grosz, Andrew I(ehler, l(aren 
Lochbaum, Stuart Shieber and Candy Sidner for helpful com- 
ments regarding tiffs paper. This researdt has I)een supported 
by U S WEST Advanced Technologies, by the Air Force Office 
of Scientific Research under Contract No.AFOSR-89-0273, and 
by an IBM Graduate Fellowshila. 
ACRES DE COLING-92, NANTES, 23-28 h0~r 1992 2 6 7 
least in part, these two challenges. As their names 
suggest, means clauses express the means by which 
an action is performed, while rationale clauses express 
the reason, or rationale, of the main clause action. In 
the last section of this paper, we show how to distin- 
guish rationale clauses from two superficially similar 
constructions: purpose clauses, as in "Mary bought a 
suit to wear at the meeting", and infinitival relative 
clauses, as in "John found the book to give to his sis- 
ter". 
(1) ~. Mary pressed the button to reset the printer. 
b. Mary reset tbe printer by pressing the button. 
Rationale clauses and means clauses are of interest 
not only as exemplars of multi-action utterances, but 
also because of the subtle differences in the information 
that can be felicitously inferred from their use. For ex- 
ample, despite the similarities between (la) and (lb), 
the continuation in (2), below, is felicitous after (la) 
but not after (lb). The interpretation task therefore 
requires determining the relations between the actions 
described in multi-action utterances and drawing ap- 
propriate inferences about them. This particular need 
is one of the issues addressed in this paper. 
(2) But that didn't work because she forgot to plug it in. 
We start by presenting distinctive properties of ut- 
terances with rationale clauses and means clauses. Af- 
ter describing their logical form, we then present the 
generation, enablement and Contributes relations be- 
tween actions \[3, 13\], arguing for their use in the in- 
terpretation of these constructions. We finally define 
interpretation rules for deriving the meaning of the 
logical form predicates representing means clause and 
rationale clauses. We also show how these rules pre- 
dict the properties of these constructions mentioned 
above. We end the paper by discussing the scope of the 
analysis and showing how the constructions to which 
these rules apply can be distinguished algorithmically. 
A companion paper \[5\] examines the mental states of 
the speaker and performing agent of such utterances, 
and describes the interpretation model. 
2 Properties of rationale and 
means clauses 
A detailed analysis of multi-action utterances in task- 
oriented dialogues \[3\] and a subsequent examination of 
a set of selections from the Associated Press news wire 
Pace. OF COLING-92, NANTES, AUG. 23-28, 1992 
revealed the following distinctive features displayed ill 
utterances with rationale clauses and means clauses. 
2.1 Semantic properties 
Overlap: In utterances with means clauses or ra- 
tionale clauses, the performance of one of the actions 
referred to in the utterance may result in the simul- 
taneous performance of the other (but need not, see 
below). The overlap in meaning between utterances 
(la) and (lb) exemplifies this property: in both utter- 
ances, Mary's pressing the button may result in her 
resetting the printer at the same time. 
Distinction: It follows from past tense utterances 
with means clauses, but not from those with rationale 
clauses, that the agent's "goal action" was success- 
fully performcd. This action is described in the main 
clause of an utterance with a means clause, but in tim 
adjunct clause of an utterance with a rationale clause. 
This property is reflected in tbe difference between ut- 
terances (la) and (lb) with respect to tile continuation 
in (2): only from (lb) can it be concluded that Mary 
successfully reset the printer 1. Thus (2) is not felic- 
itous in the context of (lb), but it is in that of (la) 
because in that context, Mary might have failed to 
reset the printer. 
Ambiguity: Utterances with rationale clauses are 
ambiguous with respect to the relative timing of the 
two actions referred to: the performance of the main 
clause action may result in the simultaneous perfor- 
mance of the adjunct clause action, or it may only 
make possible its subsequent performance. This prop- 
erty explains why, in the examples helcw, response 
(3b) is felicitous whereas the same utterance in tile 
context of (4a) is not. 
(3l a. John: I inserted a diskette to back up my file. 
b. Mary: Okay, now back up the file. 
(4) a. John: I typed 'cp new.txt' to back up my file. 
b. Mary: * Okay, now back up the file. 
2.2 "Algebraic" properties 
I)rawing an analogy with the algebraic properties of 
relations, utterances with means clauses and rationale 
clauses are irreflexive, asymmetric and transitive, a~s 
reflected in the following examples (the first utterance 
or utterance pair of each set includes a means clause, 
th,' second, a rationale clause). 
I',l a.*Mary reset tile printer by resetting the printer. 
*Mary reset the printer to reset the printer. 
b. Mary reset the printer by pressing the button. 
*Mary pressed the button by resetting the printer. 
Mary pressed the button to reset the printer. 
*Mary reset the printer to press the button. 
c. Mary fixed her problem by resetting the printer 
by pressing the button. 
Mary fixed her problem by pressing the button. 
1 The companion paper \[5\] addresses the issue of embedding 
inferences like this one in belief contexts. 
Mary pressed the button to reset the printer to 
fix her problem. 
Mary pressed the button to fix her problem. 
Similar examples may be found to show that utter- 
ances with rationale clauses about sequential actions 
have tile same properties, e.g., for transitivity: 
(6) John borrowed the book to bring it back lmlne. 
John brought the book back home to look for 
papers on adjunct clauses. 
John borrowed the book to look for papers on 
adjunct clauses. 
Transitivity affects utterances with means clauses 
and rationale clauses in different ways. The transi- 
tivity of utterances with means clauses, illustrated in 
the first utterance pair of (5c), makes it possible to de- 
scribe actions at lower levels of detail, i.e., with greater 
granularity. The transitivity of utterances with ratio- 
nale clauses, and in particular those that describe se- 
quential actions as illustrated ill (6), increases not tile 
granularity of the description, but the "distance" be- 
tween tile actions described, i.e., the number of steps 
in tile sequence of actions linking them. 
3 Actions and action relations 
in logical forms 
Logical forms represent the literal meaning of all ut- 
terance and are derived cornpositionally by seman- 
tic interpretation rules based on the syntactic struc- 
ture of tlre utterance. Our logical forms are be.sed 
on Davidson's formalism \[6\], because action reifica- 
tion facilitates the representation of utterances about 
multiple actions and relations among thent. These 
logical forms are existentially quantified sentences of 
first-order logic, with predicates that include all ad- 
ditional argument position for the action being de- 
scribed. Davidson's framework is extended for nmlti- 
action utterances by introducing multiple action vari- 
ables and by representing connectives like "hy" (intro- 
ducing means clauses) and "to" (introducing rationale 
clauses) using two-place predicates ranging over these 
action variables, as illustrated in (7) ~. 
(7) a. Mary reset the printer by pressing the button. 
LF: Sz~, x2, reset(zj ) h past(xl ) A ayt(x~, Mary) 
Aobj(xl, Printer) A by(x1, x2) 
Apress(x~) A obj(x,z, Button) 
b. Mary pressed the button to rcsct the printer. 
LF: 3xl,x2,press(x~ ) A past(xl) A agt(xl,Mary) 
Aobj(x,, Button) A i ..... der_to(z,, x~) 
^ .... t(x:) ^ obj(x2, Printer) 
The following subsections discuss the four main 
components of these representations: action variables 
2Extensional representations of propositions, like this one, 
have difficulties with respect to certain intensional phenomena 
sucb as substitution of identicals. Hobbs discusses this problem 
and a number of possible solutions Ill)l, but these questions lie 
beyond the scone of the nresent paner. 
AcrEs DE COLING-92, NANTES, 23-28 Aot~'r 1992 2 6 8 PROC. OF COLING-92, NANTES, AUG. 23-28. 1992 
(e.g., xl and x2), action descriptions (e.g., tile predi- 
cates reset and press), tense predicates (e.g., tile pred- 
icate past), and action relations (e.g., the predicates 
by and in_order_to). 
3.1 Action variables 
While Davidson's formalism assumes that action vari- 
ables denote "reM" action tokens, action variables in 
our formalism correspond to action entities in the 
discourse model. The ability to refer to an entity 
anaphorieally has been suggested as a test for "dis- 
course entity hood" \[')0\]. Actions referred to in means 
clauses and rationale clauses pass this test, as illus- 
trated in (8). The pronoml 'that' in this utterance is 
ambiguous, but the possibility of its coreferring with 
the adjunct clause action is sufficient justification for 
attributing discourse entity-hood to that action. 
(8) Mary will clone the rhinovirus to make a vaccine 
for the connnon cold and that will make her famous. 
Action entities may represent actual, or "real", ac- 
tions, i.e., actions that have occurred (e.g., the reset 
action in (7a)), or unrealized actions, i.e., actions that 
have not yet occurred or might never occur (e.g., the 
reset action in (7b) that may or may not result froln 
the press action) \[4\]. Quantification in the logical 
form is therefore over a universe that contains every- 
thing that can be talked about, with no commitment 
to existence ill the real worhl (as in, e.g., \[10, 17\]). 
Discourses in collaborative settings often involve dis- 
cussions about unrealized actions. Agents, for exam- 
pie, might make suggestions for future actions, or dis 
cuss the advantages and disadvantages of a particular 
course of action. It is thus crucial to adopt a represen- 
tation that will account for these actions as well. 
3.2 Action descriptions 
Actions arc rcprescntcd by predicates taking a single 
action variable, e.g., reset(x)), and every action ar 
gmncnt is rcprescntcd by a scparate predicate, e.g., 
agt(xl,Mary). This type of representation has been 
proposed by a number of researchers (e.g., \[21\]) as 
an alternative to Davidson's representation, in which 
only the optional arguments arc represented as sepa- 
rate predicates. One reason is the arbitrariness that is 
sometimes involved in deciding on thc argument struc- 
ture of various action predicates. For the present anal- 
ysls, this type of representation is usefld because it 
avoids having either to introduce existential variables 
for missing arguments or to determine how these ar- 
guments are recovered during the process of building 
tile logical form. Missing subject and time arguments 
of adjunct clause actions in tile logical forms of means 
clauses and rationale clauses (e.g., those in (7)) can 
he recovered through the interpretatk)n of the action 
relations (see Section 5.2). Not all rationale clauses. 
however, \[lave empty subject positions, e.g., "Mary 
prepared lobster (in order) for her gnests to have a 
taste of New Enzland cuisine." 
3.3 Tense predicates 
The tense predicates in our logical forrns capture two 
pieces of information: temporal information and in- 
formation about action occurrence. This distinction is 
necessary because actions, whether their time is past, 
present, or future, need not occur. For example, agents 
may discuss a future action without ever performing 
it. Similarly, agents ruay talk about a past action that 
never occurred, as in "John said that Mary left yester- 
clay, hut he was wrong". 
If thc logical form predicate pasl holds of some ac- 
tion, then that action occurred (in some world or belief 
context) and its occurrence was in the past. In (7), past 
is predicated of the action variable x t, denoting the ac- 
tions described in the main clauses of both (Ta) and 
(7h). If there is no information to the contrary (e.g., 
tam utterance is not embedded in an intensional con- 
text, established either directly within the utterance, 
as in "Mary thought she reset the printer", or through 
tile discourse context, as when describing a dream), 
then we call conclude that the action occurred in the 
real world. 
Utterances with rationale chmses and mcans clauses 
differ in what reformation they communicate about 
the occurrence of the action described in the adjunct 
clause. In lmithcr construction does this clause in- 
volvo a tensed verb, whidl explains why pasl is not 
predicated of x2 in the LFs of (7a) and (Tb). It is 
possible, however, to infer from (Ta) that the action 
denoted by that variable did occur, and from (7b) that 
it lnight not have occurred, ms illustrated in (1) and 
(2) m the Introduction. This information is captured 
by the interpretation rules for means clauses and ra- 
tionale clauses, as explained in Section 4. 
3.4 Action relations 
Finally, the l,Fs in (7) include predicates representing 
the connectives that link the two clauses in each of 
the utterances, namely by and in_order_to. The mean- 
lug of these predicates is specified by interpretation 
rules, presented in the next section, that state neces- 
sary conditions on the use of these constructions. It is 
important to note that these predicates are used only 
to represent the connectives "by" and "to"introdncing 
umans clauses and rationale clauses. Other uses of 
these lexical items are recognizable in a number of 
ways (e.g., syntactically: "by" followed by a noun 
phrase, or "to" introducing an infinitival complement 
(argnment) to a verb, ms in "John wanted to leave"), 
~s discussed in Section 5.3; these arc assumed to be 
represented differently in the I,F. 
4 Interpreting action relations 
Three relations between actions -- generation, enable 
mcnt and (~ontrihutes -- that have been motivated in- 
dependently by research ill plan recognition \[15, 3, 13\] 
can he employed in tile internretation of utterances 
AcrEs DE COL1NG-92, NANTES, 23-28 AO~r 1992 2 6 9 PROC, OF COLiNG-92, NANTES, AUG. 23-28. 1992 
with means clauses and rationale clauses. In this sec- 
tion we demonstrate this use of these relations, and 
also provide a definition of enablement. 
4.1 Generation 
The generation relation, initially defined by Goldman 
\[8\], holds of two actions when the performance of the 
first results in the simultaneous performance of the 
second. Generation, however, encompasses more than 
simply two simultaneous actions. For example, Mary 
can reset the printer by pressing a button only in cer- 
tain circumstances, namely when the electricity is on, 
the printer is plugged in, etc. Performing the gener- 
ating action when these conditions do not hold does 
not result in the performance of the generated action. 
Generation, therefore, is a relation that depends upon 
contextual constraints, called generation-enabling con- 
ditions \[15\], that must hold during performance time. 
This relation was formalized by Pollack \[15\], and sub- 
sequently revised in previous work of the author's \[3\]. 
The following English gloss of the definition suffices 
for the present discussion and avoids an unnecessary 
digression into an explanation of the formalization; A 
and B are actions, comprising an act-type, agent and 
time. 
(9) A generates B if and only if 
1. A and B are actions performed by the s~me 
agent at the same time 
2. there is a set of conditions, C, such that 
a. these conditions hold during the 
performance .time of A and 
b. the act-type of A conditionally generates 
the act-type of B under C. 
Clause (2b) states the requirement that there bc a 
condilional generation relation between the two act- 
types and the generation-enabling conditions, C. This 
relation abstracts away from specific actions to cap- 
turc the regularities of a generation relation across oc- 
currences. It holds between act-types ~ and fl and 
conditions C if (a) whenever an action of type c~ oc- 
curs while these conditions hold, an action of type 
occurs at the same time and (b) neither an occurrence 
of an action of type a, nor C holding, is by itself suffi- 
cient to guarantee the occurrence of an action of type 
/3 \[15\]. An important property of the generation rela- 
tion is that if it holds of two actions A and B, and A 
occurs, then so does B. This follows from the genera- 
tion relation requiring the relevant conditions to hold 
during performance time (as stated in clause (2a) in 
tile definition). 
In light of the generation relation, we can recon- 
sider the examples in (1). Both these utterances refer 
to an underlying conditional generation relation: in 
the right circumstances, pressing the button results 
in simultaneously resetting the printer. They differ, 
however, in that they do so from different perspec- 
tives: utterance (lb), with the means clause, asserts 
the occurrence of the generated action (the resetting 
action), while utterance (la). with the rationale clause. 
does not (cf. example (2)). Therefore, tbe relevant 
generation-enabling conditions must have held during 
performance time in the context of (lb) bug not neces- 
sarily in that of (la). Because the generation relation 
requires these conditions to hold during performance 
time, it correctly describes the relation between the 
two actions in (lb) but not necessarily that in (la). 
Utterances with rationale clauses, like (la), only 
have the potential for realizing a generation relation, 
and this potentiality rests on two different factors: (a) 
the possibility that there is no conditional generation 
relation between the two actions, as in "Mary ate an 
apple to reset the printer", or (b) the possibility that 
tbe generation-enabling conditions do not hold during 
performance time, as in (la) where the two actions are 
related by a conditional generation relation, but the 
associated generation-enabling conditions might not 
hold during performance time. 
The following sections, however, will show that an 
analysis of rationale clauses based solely in terms of 
the generation relation is too restrictive. 
4.2 Enablement 
The interpretation of rationale clauses also requires ap- 
pealing to the enablemen~ relation; for example, in (10) 
below (of. (la)), tile performance of the main clause 
action does not (potentially) result in the simultane- 
ous performance of the adjunct clause actiou. Rather, 
it only makes possible the subsequent performance of 
the adjunct clause action. The generation relation, 
tberefore, does not correctly describe the action re- 
lation in this utterance. Instead, tile relationship is 
oue of enablement, a relation we define ms holding of 
two actions when the performance of tile first brings 
about a condition that is necessary for the subsequent 
performance of the second \[3\]. 
(10) Mary inserted a diskette to back up her file. 
Much of the literature that makes use of tile enable- 
ment relation has been vague about its intended inter- 
pretation, using it at different times to mean different 
things and not explaining wily one action nmst be per- 
formed before the other (e.g., \[18\]). In the following 
definition, the dependency between the enabling and 
enabled actions is explained by the role of an interven- 
ing conditional generation relation 3. A closer exami- 
nation of utterance (10), for example, shows that the 
effect of the enabling action, namely that a diskette be 
in place, provides the generation-enabling condition of 
a conditional generation relation between typing somc 
command and backing up thc associated file. 
(11) A enables B if and only if 
1. the time of A is prior to the time of B, 
2. there are a set of conditions, C, and a third 
action A' such that 
a. one of the eonditious in C, Ci, holds 
as a result of the performance of A, and 
b. the act-type of A' conditionally generates 
the act-type of B under C. 
3This is not the only oossible type of delpendellcv \[3\], 
AcI~ DE COLING-92, NANTES, 23-28 AOtTtr 1992 2 7 0 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
If Ci, the condition brought about by the perfor- 
mance of A, is the only element of the set of generation- 
enabling conditions C, then the performance of A is 
both necessary and sufficient for the subsequent I)cr- 
formance of B. 
4.3 Contributes 
So far, utterances with rationale clauses have been in- 
terpreted as instances of the generation or enablemcnt 
relation. This construction, however, can also relate 
two actions when the chain of relations between them 
involves both generation and enablement relations, as 
illustrated in (12a), or a complex action constructor 
\[3\], as illustrated in (12b): 
(12) a. Mary built the table to sell it at the market. 
Mary sold the table at the market to earn money. 
Mary earned money to pay her rent. 
Mary built the table to pay her rent. 
(enablement + generation + enablement) 
b. l picked up the foot end for us to lift the piano. 
(simult + generation) 
Utterances with rationale clauses thus allow for a 
wide range of relations between the main clause and 
adjunct clause actions. To capture this phenomenon, 
we interpret them as iustances of the Contributes rela- 
tion. This relation, also originally introduced for plan 
recognition purposes, is defined as the transitive clo- 
sure of the generation and enablement relations, and 
the complex action constructors \[13\]. It therefore cov- 
ers the entire range of relations that can exist between 
the actions in this type of utterance. 
5 Analysis 
5.1 Interpretation rules 
The inference rules defining the meaning of means 
clauses and rationale clauses arc given below. Tile 
by and in_order_to predicates in the left hand side of 
the rules are the logical form predicates representing 
the connectives introducing these clauses 4. Tbe verbs 
"generates" and "contributes" in the right hand side 
are used here, as elsewhere in tbe paper, in their tech- 
nical sense, corresponding to the action relations de- 
fined in the previous sections. Note that the order 
of the arguments in the by predicate and generation 
relation is reversed. 
(13) a. LFI: by(h,B)-~ 
B generates A and B occurred. 
b. LF2: ia_order_to(A,B)---* 
A contributes to B, 
within restricted belief contexts. 
The LF1 rule maps the LF representation of means 
clause to the generation relation and specifies that the 
4This paper does not address the reference problem and as- 
sumes that the mapping of action variables in the logical form to 
action entities in the discourse model has already been aonlied. 
generating action occurred. This occurrence is a fea- 
ture of the means clause construction and not of the 
use of the gerund form to describe the action. Given 
an utterance with a means clause, it can therefore be 
inferred that the generated action occurred as a result 
of the generating one and that the relevant contextual 
conditions held during performance time. 
The LF2 rule maps the LF' representation of ratio- 
hale clauses to a Contributes relation within a belief 
context, capturing the fact that (a) the relation can 
be any combination of the generation and enablement 
relations, as well ms the complex action constructors 
and (b) that the relation between tile two actions is 
only potential. A new set of action relatious is not 
needed for representing the potentiality of the action 
relation in these utterances. Instead, as described in 
the companion paper \[5\], it is best expressed in terms 
of the different beliefs and intentions of the speaker 
(S) and performing agent (G). For example, G could 
believe a Contributes relation between actions A and 
B, wlfite S does not; or G and S could both believe 
the relation, but haw~ different beliefs regarding asso- 
ciated conditions and actions, e.g., G may believe A is 
necessary and sufficient for B whereas S believes A is 
only necessary. 
When utterances with rationale clauses involve si- 
lmlltancous actions, then the Contributes relation de- 
rived by the LF2 rule specializes to the generation re- 
lation. Exactly when such specialization is possible is 
a plan recognition problem that goes beyond the scope 
of this paper. Utterances sometimes provide tempo- 
ral or agency information that can be used to opti- 
mize the plan recognition process. In particular, if the 
performance times, or performing agents, of the two 
actions are known to bc different, then a specializa- 
tion to generation is ruled out. When temporal and 
agency information is not available, then recipes pro- 
vide anotber source of information for disambiguating 
rationale clauses \[3, 14\]. 
5.2 Consequences of the rules 
The two interpretation rules LF1 and LF2 predict all 
of the properties of utterances with rationale clauses 
and means clauses that were presented in Section 2. 
Semantic properties: The LF2 rule accounts for 
the Ambiguity property of rationale clauses by making 
use of the Contributes relation which can relate either 
simultaneous or sequential actions. Because the Con- 
tributes relation may specialize to a generation rela- 
tion, the LFI and LF2 two rules account for the Over- 
lap property. Finally, because tile LF2 rule is based 
on potentiality, while tbe LF1 rule is not, these rnles 
also account for tile Distinction property. 
Algebraic properties: It can be shown that the 
generation and Contributes relations are irreflexive, 
asymmetric and transitive, rellecting the algebraic 
properties of the NL utterances that realize them. 
By describing simultaneous actions performed by the 
same agent, the generation relation also captures the 
increased ~rannlaritv of the action description: bv de- 
AcrEs DE COLING-92, NANTES, 23-28 AOt3T 1992 2 7 1 PROC. OF COLING-92, NANTES, AUo. 23-28, 1992 
scribing sequential (as well as simultaneous) actions, 
the Contributes relation captures the increased dis- 
tance between the actions described. 
In addition, these interpretation rules account for 
the missing agent and time arguments of the actions 
described in means clauses and rationale clauses. Be- 
cause the generation relation requires the agents and 
times of the two actions to be identical, the miss- 
ing arguments in a means clause can be recovered on 
the basis of the agent and time information associ- 
ate'd with the main clause action. The Contributes 
relation, however, imposes no temporal or agency con- 
straint. Therefore, without additional information, it 
is not possible to determine whether the agent (when 
missing) and time of a rationale clause are the same as 
those of the main clause, el. "I brought this miserable 
Morgon to enjoy with our dinner" \[2\]. 
5.3 Scope of the analysis 
The LF1 and LF2 interpretation rules predict specific 
facts about English, namely a large set of properties 
characterizing the differences and similarities between 
rationale clauses and means clauses. They do so by 
defining the meanings of the by and in_order_to LF 
predicates in terms of independently motivated action 
relations. These predicates are useful for representing 
not only the meaning of rationale clauses and means 
clauses, but also that of related constructions. For 
example, tile in_order_to predicate can be used to rep- 
resent the connective "so that" in utterances such as 
"Mary got up early so that she could catch the first 
train to Montreal". The by predicate can also be used 
for non-clausal constructions, as in "John got the infor- 
mation he needed by means of an interview". Further- 
more, these predicates are meant to be language inde- 
pendent. In French, for example, the in_order_to predi- 
cate would be used to represent the construction "pour 
+ infinitive", as in "Marie a regard~ des films pour 
preparer son voyage au Japon" (Mary watched movies 
to prepare for her trip to Japan), and the by predi- 
cate, to represent the construction "en + gerund", as 
m "Marie a pr~par~ son voyage au J upon en regardant 
des films" (Mary prepared for her trip to Japan by 
watching movies). Unlike the corresponding English 
construction, however, the French "en-clause" is am- 
biguous between a means clause reading (as given in 
tile English translation) and a temporal reading that 
can be translated as "while watching movies". Deter- 
mining how to disambiguate these two interpretations 
is an interesting research question. 
In the remainder of this section, we show how to 
distinguish algorithmically the constructions to which 
the LF1 and LF2 rules apply, thereby making it pos- 
sible for an interpretation system to build the correct 
LF representations. 
Rationale clauses: These clauses are optional in- 
finitival adjuncts, introduced either by the infinitival 
marker "to", or, when the subject is not missing, by 
the lexieal item "for" followed by a noun phrase (re- 
alizinz the subiect\] followed by "to". Both these cola- 
structions are optionally preceded by "ill order". Note 
that the LF2 rule applies to rationale clauses with or 
without explicit subjects. For example, given "Mary 
prepared lobster (in order) for her guests to have a 
taste of New England cuisine", Mary's preparing lob- 
ster contributes to her guests having a taste of New 
England cuisine. 
A large number of to-infinitival clauses (with or 
without subjects) to which the LF2 rule does not ap- 
ply can be distinguished on syntactic grounds, namely 
by not being optional infinitival adjuncts. Utterances 
that are syntactically ambiguous show a correspond- 
ing semantic ambiguity, and the LF2 rule applies to 
the rationale clause reading; e.g.: "Bush warned his 
press secretary to prevent a leak", where tlle infiniti- 
val clause may function as an optional adjunct (i.e., 
realizing a rationale clause) or a complement. 
Two types of adverbials to which the LF2 rule dees 
not apply, but which are syntactically indistinguish- 
able from rationale clauses, are comment clauses, as 
in "John is not sure what to do, to put it brietly", and 
outcome clauses, as in "Mary awoke one morning to 
find the house in an uproar" \[16\]. Given their status 
as parentheticals, comment clauses are recognizable in 
text by a comma, and in speecb by intonation (note 
that without the comma, the infinitival clause becomes 
a rationale clause). Outcome clauses are restricted to 
verbs of discovery or to the verbs "be" or "become" 
(e.g., "She lived to be 10ft'), and are therefore also 
distinguishable. 
Finally, there are pnrpose clauses, e.g., "Mary 
bought a suit to wear at the meeting". Purpose clauses 
are also optional infinitival adverbials, but differ from 
rationale clauses in the mnnber and type of gaps they 
allow and in their control pattern s \[2, 12\]. They also 
differ from the superficially similar infinitival relative 
clauses by not being attached to a noun phrase. Tile 
LF2 rule, in fact, makes correct predictious for this 
construction as well, e.g., Mary's buying a suit con- 
tributes to her wearing the suit at tile meeting. Pur- 
pose clauses with inanimate subjects, however, as in 
"John bought a shelf to hold his books", require some 
adjustment with respect to tbe action's role in the 
Contributes relation. Actions relations are between ac- 
tions, and actions are assumed to have human agents. 
Therefore, ratber tban deriving a Contributes relation 
between John's buying a shelf and the shelf's hold- 
ing his books, the system should derive a relation be- 
tween John's buying a shelf and his having something 
(namely the shelf) to hold his books. This issue is 
currently being investigated. 
Means clauses: These clauses are very easily dis- 
tinguishable by their form: the lexical item "by" fol- 
lowed by an -lug clanse (in this context, "by" may be 
treated as a preposition or a subordinator). Appar- 
ent counterexamples to the LF1 rule are utterances of 
the form "Mary pays her rent by building filrniture", 
5Fl'om a descriptive point of view, rationale clauses differ 
fro~. purpose clauses in that only the former Mlow the para- 
nhrnae with the words "in order to" \[121. 
AcIXs DE COLING-92, NANTES, 23-28 ^o',rr 1992 2 7 2 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
in which the two actions are sequential. However, the 
fact that LF1 does not apparently apply to this ex- 
ample, is a sign that additional pragnmtic reasoning 
(e.g., conversational implicatures \[9\]) is needed. In 
this particular example, LF1 does apply, though indi- 
rectly through the utterance of which it is a metaphor, 
namely, "Mary makes a living by building furniture". 
The need for further Gricean reasoning of this sort is 
reflected in the markedness of such examples. 
6 Conclusion 
In this paper, we analyzed utterances with ratio- 
nalc clauses and means clauses, presented their log- 
ical form, and defined interpretation rules for these 
two constructions that make use of independently mo- 
tivated action relations. These rules make it possible 
to draw appropriate inferences about the actions de- 
scribed in these utterances, thereby accounting for thc 
similarities and differences between rationale clauses 
and means clauses. The algebraic properties of tile ac 
tion relations were shown to reflect the properties of 
the utterances that realize them. We also showed how 
to distinguish thc constrnctions to which thesc rules 
apply. 
This analysis fills an important gap in the natural 
language interpretation literature. Surprisingly little 
research has examined multi-action utterances such 
as those with means classes and rationale claases. 
There is a large body of linguistics research on pur- 
pose clauses, but it focuses on syntactic aspects of the 
construction and issues of control (e.g., \[2, 12\]). In the 
computational linguistics literature, Huettner et al's 
work on the generation of adjunct clauses \[11\] and Di 
Eugenio's analysis of instructions \[7, 19\] iloth examine 
purpose constructions, but from different perspectives 
than that of this paper \[5\]. 
This research is being extended ill two main direc- 
tlons. One is to examine future tense utterances and 
other types of utterances about unrealized actions. 
Another is to cast the interpretation rules in a genera\[ 
processing model that takes into account the mental 
states of the agents involved. Beliefs and intentions arc 
necessary for the treatment of unrealized actions, ms 
well as for capturing other aspects of rationale clauses 
and means clauses. In particular, utterances with ra- 
tionale clauses, but not necessarily those with means 
clauses, communicate the speaker's beliefs about the 
performing agent's intention to perform the actions 
described in the utterance. Finally, mental states arc 
indispensable in any interpretation model that pur- 
ports to account for inferences that arc drawn on the 
basis of utterances, initial progress in these two areas 
is reported in tile companion paper \[5\]. 
ACTES DE COLING-92, NANTES, 23-28 AOUT 1992 2 7 3 PROC. OF COLING-92, NANTES. AUG. 23-28, 1992 

References 

\[1\] Albert, D. 1990. Cooperative dialogs while play 
ing adventure. TR-O2-90, llarvard Univ. 

\[2\] Bach, E. 1982. Purpose clauses and control. Ill 
Jacobson, P. & t'ullom, G., ells, The Nature of 
Syntactic Representation. 1). lteidel l'ubl. (7o. 

\[3\] Balkanski, C. 1990. Modelling act-type relations 
in collaborative activity. Technical Report TRo 
23-90, Ilarvar(I Univ. 

\[4\] Balkanski, C. 1991. Logical form of complex ac- 
tion sentences in ta.sk-orimtted dialogs. In P1vc. 
ACL'91, Student Session. 

\[5\] Balkanski, C. 1992. Actions, beliefs and inten- 
tioas in rationale clauses and means claases. In 
f'roc. AAAU92. 

\[6\] I)avidson, D. 1967. Thelogical lbrmofaction sen- 
tences, ht l~.escher, N., ed, The Logic of Decision 
and Action. University of Pittsburgh Press. 

\[7\] Di Eugenio, 1t. 1992. Goals and action in natural 
language instructions. Technical Report MS-CIS- 
92-07, Univ. of Pennsylvania. 

\[8\] (;oldrnan, A. 1970. A Theory of Human Action. 
Princeton University Press, Princeton, NJ. 

\[9\] (\]rice, \[I.P. 1975. Logic and conversation. In Syn- 
tax ~4 b'emanlics, vol 3, Speech Acts, Academic 
Press, NY. 

\[10\] Ilobbs, J. 1985. Ontological promiscuity. In PTve. 
A CL '85. 

\[11\] lluettner, A., Vaughan, M., and McDonald, D. 
1987. Constraints on the generation of adjunct 
clauses. In Proc. ACL'87. 

\[12\] Jones, C. 1991. l'urpose clauses: syntaz, themal- 
ics and semantics of English purpose construe- 
lions. Kluwer Academic Publishers. 

\[13\] l,ochbaum, K., Grosz, B., and Sidner, G. 1990. 
Models of plans to support communication: An 
initial report, lu Proc. AAAI'90. 

\[14\] Lochbamn, K. 1991. An algorithm for plan 
recognition in eollaboratiw~ discourse. In Proc. 
ACL '9I. 

\[15\] Pollack, M. 1986. Inferring domain plans in 
question-answering. Technical Report 403, SRI 
International, Menlo Park, CA. 

\[16\] Qnirk, R., Greenbaum, S., Leech, G., and 
Svartvik, J. 1985. A Comprehensive Grammar 
of the English Language. Longman. 

\[17\] Schubert, L. K. and Hwang, C. tl. 1990. An 
ci)isodic knowledge representation for narrative 
texts. Tcdmical Report 345, Univ. of Rochester. 

\[18\] Schuster, E. 1988. Anaphorie reference to events 
and actions: a representation and its advantages. 
In Proc. COLING'88. 

\[19\] Webber, B. and Di Eugenio, B. 1990. Free ad- 
juncts in natural language instructions. In PTvc. 
COLING '90. 

\[20\] Webber, B. 1983. So what can we talk about 
now? In Brady, M., and Berwick, 1L C., eds, 
Computalional Models of Discourse. MIT Press. 

\[21\] Wilensky, I(. 1991. The ontology and representa- 
tion of situations. In Proceedinys of KR'91. 
