Recognizing Digressive Questions During Interactive Generation * 
Susan M. Hailer 
Department of Computer Science and Engineering 
University of Wisconsin - Parkside 
Kenosha, Wisconsin 53141 
haller©cs, bufgal o. edu 
i 
Abstract 
In expository discourse, people sometimes ask ques- 
tions that digress from the purpose of the discussion. 
A system that provides interactive explanations and 
advice must be able to distinguish pertinent questions 
from questions that digress. It must also be able .to 
recognize questions that are incoherent. These types of 
questions require different treatment. Pertinent ques- 
tions must be answered to achieve the discourse ~ phr~ ~ 
pose. If the user asks a digressive question, the sys-• 
tem may need to shift the focus of the discussion back ' 
to the original-purpose. Incoherent questions signal a 
more serious misunderstanding that requires clarifica- 
tion and repair. 
The Interactive Discourse Planner (IDP) is designed 
to plan text to describe and/or justify a domain plan 
interactively. As a testbed, IDP plans text to discuss 
driving routes. IDP uses questions from the user to 
recognize how to extend its own text plan in a way 
that both satisfies its listener and achieves the system's 
discourse goal. In the process of recognizing ways to 
expand its own text plan, IDP can detect three types of 
digressions that the user can initiate with a question. 
1 Characterizing Digressions 
Grosz and Sidner define a digression as a type of :inter- 
ruption \[4\]. An interruption is a discourse segment with 
a purpose that does not contribute to the achievement of 
the current discourse purpose. They describe three kinds. 
A true interruption has a discourse purpose that is unre- 
lated to the interrupted discourse segment. For example, if 
a speaker says 
"John came by and dropped off the groceries Stop that 
you kids. and I put them away after he left." 
the italized portion of text is a true interruption. 
Speakers use a second type of interruption, the flashback 
or filling ~n missing places to bring objects and propositions 
into the discussion that aid in comprehension of the current 
discourse segment. This type of interruption provides back- 
ground knowledge. However, it does not contribute directly 
to the current discourse purpose. For example in 
*I would like to thank my advisor, Stuart C. Shapiro, and the 
members of the SNePS Research Group in the Computer Science 
Department at the State University of New York at Buffalo. 
Their advice and comments are reflected in the research that 
this paper describes. 
"OK. Now how do I say that Bill is Whoops I forgot 
about ABC. I need an individual concept for the com- 
pany ABC... Now back to Bill. How do I say that Bill 
is an employee of ABC?" 
the speaker sets aside her current purpose to discuss a pre- 
•. requisite that should have been introduced earlier. 
We are concerned with a third type of interruption that 
Grosz and Sidner describe as a digression. A digression con- 
• tams a reference to some entity that is salient in both the 
interruption and the interrupted segment. The digression's 
PUrpose is not unrelated to the purpose of the interrupted 
segment (as in a true interruption), but neither is it a pre- 
requisite (as in a flashback). As an example of a digression, 
Grosz and Sidner note that if while discussing Bill's role in 
company ABC, a conversational participant interrupts with 
Speaking of Bill, that reminds me, he came to dinner 
last week, 
Bill remains salient, but the discourse purpose changes. 
In the IDP model, the system's purpose (or intent) is 
expressed by its discourse goal and the text plan that the 
system formulates and executes to try to achieve it. IDP 
detects a digression when the user asks a question about 
a discourse entity or a proposition that is part of the sys- 
tem's text plan, and the answer to the question cannot be 
ir/corporated into the text plan. 
IDP Operates in a cooperative interactive mode in which 
the System is the primary speaker and the user is the pri- 
mary listener. Henceforth, I refer to the user as the lis- 
tener. In this interactive mode, IDP controls the discussion 
to make its own intentions clear and to try to achieve them. 
2 Digression Types 
The first and simplest kind of digression that IDP can de- 
tect is when the listener asks a direct question about a 
proposition that was expressed by the system earlier in the 
discourse. The following interaction demonstrates this kind 
of digression and how the system responds to it. 1 
i ¸ . 
1All demonstrations given in courier font are actual inter- 
actions with IDP. 
181 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
Demonstration I: 
L: Should I take Maple or Sheridan to go to 
the Eastern Hills Mall? 
IDP: if now there was light traffic, 
you could take Sheridan however, 
now there is heavy traffic. 
you should take Maple. 
since taking Maple there are fewer 
businesses than taking Sheridan, 
taking Maple avoids heavy traffic. 
L: Why is there heavy traffic now? * 
IDP: since now is rush hour, 
as I was saying taking Maple avoids 
heavy traffic. 
As part of IDP's advice, it has expressed a proposition 
that lets the listener know circumstances that motivate the 
advice. In particular, IDP has told the listener that now 
there is heavy traf:fic. The system identifies the lis- 
tener's question about this statement (marked with an as- 
terisk) as a digression from the discussion purpose. There- 
fore, after answering the question, IDP shifts the discussion 
back to its recommendation. 
A second kind of digression that IDP can detect is when 
the listener questions a proposition that was never ex- 
pressed, but that can be inferred from a proposition that 
has been presented. The following demonstration illustrates 
this digression type. 
~he listener's line of questioning keeps referring to the last 
proposition that was expressed. This eventually leads to an 
explanation that does not address IDP's original purpose. 
The following interaction illustrates digressions of this kind. 
Demonstration 3: 
L: Should I take Maple or Sheridan to go to 
the Eastern Hills Mall7 
IDP: take Maple. 
L: Why7 
IDP: taking Maple avoids heavy traffic. 
L: Why? 
IDP: taking Maple there are fewer 
businesses than taking Sheridan. 
L: Why? * 
IDP: since taking Maple is a newer route 
than taking Sheridan. 
anyways, taking Maple avoids heavy 
traffic. 
Unlike the first two questions, the listener's third why- 
question requires an answer that cannot be incorporated 
into the IDP's text plan, a plan for trying to convince the 
listener to use the Maple Road route. Therefore, the system 
detects a digression. 
3 Related Work 
Demonstration 2: 
L: Should I take Maple or Sheridan to go to 
the Eastern Hills Mall7 
IDP: take Maple. 
L: Why take Maple? 
IDP: taking Maple avoids heavy traffic. 
L: Why? 
IDP: since taking Maple there are fewer 
businesses than taking Sheridan. 
L: Why is there heavy traffic now? * 
IDP: since now is rush hour, 
as I was saying taking Maple avoids 
heavy traffic. 
IDP uses its text plan to detect that the listener's third 
question, unlike the listener's first two questions, initiates a 
digression. This is notwithstanding the fact that the system 
never expressed the proposition that the listener questions. 
The listener's digressive question relies on an inference that 
she made from the proposition that IDP conveyed in its 
second response. 
IDP does not analyze every question as a digression. 
Demonstration 2a: 
L: Should I take Maple or Sheridan to go to 
the Eastern Hills Mall? 
IDP: take Maple. 
L: Why is there heavy-traffic now? * 
IDP Huh? 
In this example, IDP fails to make any coherent connection 
between the question and the text plan that they system 
has used so far. Note that the answer to this question does 
not fall outside the system's expertise. In Demonstration 2, 
IDP analyzed this question as a digression, and responded 
to it. 
A third type of digression that the system detects is when 
Moore and Swartout note that prior to their work on the 
EES Text Planner, little work had been done on using a 
sYstem's goals and plans as a model of the discourse and 
its purpose \[11, 12\]. The EES Text Planner records its 
discourse goals and the plans that it formulates to achieve 
them in a dialogue history. When the user asks a follow-up 
question, the system uses the goal structure of its text plan, 
assumptions that have been made during generation, and 
alternative plans that it did not use to disambiguate the 
question and select a perspective for the answer to it. Then 
the system formulates the appropriate discourse goal and 
passes it to the text planner to formulate a response. 
The dialogue history is a stack of text plans from previ- 
ous exchanges. Each text plan in the stack has a pointer to 
a goal (called the global contex 0 that the text plan has been 
constructed to achieve. However, the stack ordering is the 
only relationship among the text plans and their respective 
goals. There is no representation of the overarching goal 
for the interaction nor how the plan for each exchange con- 
tributes to it. Therefore, the system cannot detect when the 
user's follow-up question results in planning for a discourse 
goal that does not contribute to the original discussion pur- 
pose. 
The IDP model treats the system's text plan as a richer 
source of information. To produce answers to follow-up 
questions, IDP replans or expands a single text plan that 
has an overarching discourse goal. In this respect, IDP is 
similar to the Explanatory Discourse Generator (EDGE) 
\[2\]. EDGE plans tutorials by formulating a single text plan 
in advance and then executing it incrementally. EDGE uses 
feedback from the listener to monitor the success of its plan 
thereby making decisions to prune or extend parts of it that 
have not been used yet. 
Planning text ahead is appropriate when a system plans 
tutorials. In this situation the system has a structure of 
182 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
preset goals for the interaction. In contrast, IDP is designed 
for domains where there is no agenda for what the listener 
will come to know as a result of the interaction. IDP plans 
text to give advice and answer questions until the listener is 
satisfied. Therefore, like the EES Text Planner, IDP plans 
only in reaction to foll0w-up questions. However, to keep 
track of where the interaction is going, each plan that IDP 
constructs extends a single text plan that is formulated to 
achieve an overarching discourse goal. 
4 Text Plan Operators 
4.1 The Planning Formalism 
IDP's text plan operators (TP-operators) are based on 
Rhetorical Structure Theory (RST) \[9\] and are written us- 
ing the SNePS Actor planning formalism \[16\]. In RST, 
each essential text message (called the nucleus) can be aug- 
mented with additional information (called the satellite) 
through a rhetorical relation. In the SNePS Actor plan- 
ning formalism, plan operators are written as rules that 
state what consequent propositions can be deduced from a 
set of antecedents. 
Figure 1 shows a TP-operator for the motivate act. ~ In 
the formalism, an act decomposes into one or more struc- 
tures of other acts called plans. IDP deduces propositions 
that state how acts decompose into plans (ACT-PLAN case 
frame) by satisfying a rule's antecedents. These are the 
constraints on the plan, and the process of constraint satis- 
faction selects new content for the text. For TP-operators 
that are based on rhetorical relations, this new content is 
a satellite proposition that is appropriate to the rhetori- 
cal relation and the given nuclear proposition. The same 
constraints are used to write rules for deducing the precon- 
ditions of the motivate act (ACT-PRECONDITION case 
frame), the effects of the act (ACT-EFFECT case frame), 
and the goals that the act (viewed as a high-level plan) can 
be used to achieve (GOAL-PLAN case frame). 
The TP-operator in Figure 1 is an asserted rule (indicated 
by !) stating that for all (FORALL) ?gl, ?g2, and ?p, if the 
antecedents (ANT) can be satisfied, then the consequent 
proposition (CQ) can be deduced. The TP-operators are 
domain-independent because the constraints on them are 
stated in terms of the planning formalism. Hence, IDP's 
TP-operators are for producing text about a set of domain 
plans that are under discussion. 
In Figure 1, the constraints are that ?gl must be a goal- 
act, act ?gl must be enacted by a plan ?p, ?g2 must be a 
secondary goal-act, and act ?g2 must also be enacted by 
plan ?p. If an instantiati0n of ?gl, ?g2, and ?p satisfies the 
constraints, then the system can deduce a plan for the act 
of motivating the user to do ?p. This plan is a sequence 
(snsequence) of four other acts. The snsequence act is one 
of six control acts that are used to structure several acts 
into a plan for an act. 
4.2 The Kinds of Text Plans 
IDP uses two types of text plans (TPs) separating those 
plans that address the system's discourse goat (DG) di- 
rectly, from those plans that provide additional information 
that augments the system's essential text message. The two 
2Arguments enclosed in braces,{...}, are unordered set 
arguments. 
!FOKALL-ANT-CQ({?gI,?g2,?p}, {GOAL-ACT(?gl), 
ACT-PLAN(?gI,?p), SECONDAB.Y-GOAL-ACT(?g2), 
ACT-PLAN(?g2,?p)}, 
ACT.PLAN(motivate(user,DO(us,r,?p)), snsequenee( 
advise(user,DO(user,?p)), circumstantiate(ACT-PLAN(?g2,?p)), 
$ay(ACT-PLAN(?g2,?p)), restate(ACT-PLAN(?g2,Zp))))) 
Figure 1: A TP-operator for Motivate 
kinds of TPs are discourse text plans (DTPs) and content- 
selection text plans (CTPs). The overarching plan is always 
a DTP. This is consistent with Moore and Pollack's con- 
tention that a speaker always structures information in a 
discourse with an high-level intention in mind \[14\]. 
This division is based on a two-way division of the rhetor- 
ical relations that Mann and Thompson describe and that 
has been used by Haller \[5, 6\] and Moore and Paris \[13\] 
to design systems with two types of text plans. Each pre- 
sentational relation relates two text spans for the purpose 
of increasing an inclination in the hearer. In contrast, a 
subject-matter relation is used with the intent of informing 
the hearer of the rhetorical relation itseI£ In the IDP model, 
DTPs are used to attempt and reattempt the achievement 
of the system's discourse goals (DGs). Since these goals 
have to do with affecting the listener's attitudes and abili- 
ties towards domain plans, DTPs are based on speech acts 
and presentational rhetorical relations. The DTPs describe 
how to try to achieve DGs by selecting some minimal text 
content. IDP can augment this content without detracting 
from its own intent by using one or more CTPs. Therefore, 
CTPs correspond to subject-matter rhetorical relations. 
Figure 2(a) shows a DTP for motivate that IDP instanti- 
ates from the TP-operator given in Figure 1.3 The motivate 
act takes the listener and a nuclear clause (the listener DO- 
ing the *Maple-plan) as its arguments. The DTP for moti- 
vating the listener to take the Maple Road route sequences 
up to four other acts. The first act in the sequence, advise, 
is another DTP that must be expanded if the system has 
not already used it. The third act, say, is the only primitive 
act. When the system executes it, a text message is sent to 
IDP's surface generation grammar. The second and fourth 
acts, circumstantiate and restate, are references to CTPs 
that are potential plan growth points. Growth points are 
references to CTPs that are embedded in the body of each 
TP along the active path. Growth points suggest ways of 
adding content that augment the current TP \[7\]. 
A plan for circumstantiate is given in Figure 2(b). Since 
CTPs are not executed to affect the listener in any way 
other than to provide information, the listener is not an 
argument to acts for CTPs. IDP can only deduce a CTP 
for an act when there is an active content-goal (CG) that 
the plan satisfies. A constraint on all CTP-operators re- 
quires there to be an active CG to let the listener know the 
proposition that will be the satellite in a subject-matter 
rhetorical relation. 
As shown in Figure 2(b) as the second step in executing 
S'Maple-plan represents a plan structure stated in terms of 
going and turning acts that is not shown here. 
183 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
!ACT-PLAN(motivate(listener, D O(listener,*Maple-plan)), 
snsequence( advise(list ener, 
DO(listener,~Maple-plan)), 
cireumst antiate(ACT-PLAN(avoid(heavy-t rattic,  Maple-plan))), 
say(ACT- PLAN(avoid(heavy-t rafS¢, =Maple-plan))), 
rest ate(ACT-PLAN(avoid(heavy.t raffle, *Maple-plan))))) 
(a) 
!ACT-PLAN(cireurast antiat ¢(ACT-PLAN(avoid(heavy-t raffic, SMaple-plan))), 
snsequence( say(OB J- PR0 P(W'Maple-plan, 
fewer-businesses)), 
disbelieve(CONTENT-GOAL( KNOW(listener, 
OB J-PB.OP(~Maple.plan, fewer-businesses)), 
(b) 
Figure 2: (a) A DTP for Motivate (b) A CTP for Circum- 
stantiate 
the circumstantiate CTP, the CG is retracted (disbelieved 
by the system). Because the ACT-PLAN proposition is de- 
ducible only when the CG exists, the SNePS Belief Revision 
component (SNeBR) \[10\] retracts the ACT-PLAN propo- 
sition from the knowledge base as part of the execution of 
the CTP. Unless there is an active CG the CTP cannot be 
deduced. This keeps the CTP from being expanded and 
used even though it appears in the body of DTPs like the 
one for the motivate act (Figure 2(a)). 
5 The Analyzer 
5.1 The Discourse Context 
To make the several sources of information that are needed 
for highly interactive explanations available, I represent 
them uniformly using the SNePS Semantic Network Pro- 
cessing and Reasoning System \[15\]. IDP's DGs are un- 
achieved system goals that have to do with the attitude or 
abilities of the listener toward a domain plan. In the role 
of the primary speaker, IDP can post one or both of the 
following DGs: 
1. to have the listener adopt a domain plan 
2. to have the listener be able to execute a domain plan 
IDP plans text to try to achieve its DG, and it interprets the 
listener's feedback in the context of three types of knowl- 
edge that are all related to its TP: 
1. the active path 
2. growth points 
3. the localized unknowns 
Following Carberry, once the listener knows the system's 
intention and how it has been realized, she has expecta- 
tions for what will follow \[1\]. Motivated by Grice's Maxim 
of Relation \[3\], IDP analyzes questions using growth points 
on the active path. The active path marks the TP that 
make up the most recent expansion of IDP's overall TP. 
IDP also uses a set of propositions called the localized un- 
knowns. The localized unknowns are propositions about 
domain plans and domain-related reasoning that, based on 
the model of the listener, the listener does not know. The 
localized unknowns are linked to the system's TP by the 
reasoning chains that the system used to derive it TP. 
5.2 The Local Topic 
In the IDP model, questions address the system's DG indi- 
rectly by making reference to what the system is doing to 
achieve its DG. Therefore, IDP's Analyzer uses feedback to 
try to recognize a TP to use to expand or replan its cur- 
rent TP. Following van Kuppevelt, the local topic is that 
which is questioned, and the comment on the local topic is 
an answer to the question \[18\]. For each question, IDP's 
Analyzer determines the local topic and then searches for a 
way to expand its TP to include a comment on it. 
The Analyzer processes questions in one of the following 
forms: 
1. Why {not}? 
2. Why {not} plan? 
3. Why {not} proposffion? 
{not} indicates that the word "not" is an optional con- 
stituent of the input string. The local topic is a domain 
plan or a proposition. When there is a simple why-question 
(1), the local topic is the last proposition that was expressed 
by the system with a say-act. If the question is in the form 
of 2 or 3, IDP makes the plan or proposition that is men- 
tioned the local topic. 
5.3 A Measure of Coherence 
From the generation perspective, a conversation is a com- 
municative process that the system controls for its own pur- 
pose. In this context, the appropriate measure of coherence 
is the degree to which a new TP-expansion contributes to 
the achievement of the system's goal. Discourse expecta- 
tion constrains IDP's choices to the growth points in the 
DTPs along the active path. An important heuristic in se- 
lecting a TP-expansion is the degree to which the proposed 
expansion highlights the system's intent as realized by the 
DTP-level portion of its TP. Therefore, IDP considers the 
growth points for the DTPs on the active path in the fol- 
lowing order: 
1. DTPs that replan a DTP 
2. CTPs that expand a DTP 
IDP prefers DTPs that replan a DTP over CTPs that ex- 
pand a DTP. This encodes a preference for plans that high- 
light the system's intent over plans that supply additional 
information. 
In the first phase, IDP analyzes why-questions in rela- 
tion to its own intent as represented by the DTPs along the 
active path. The Analyzer starts with the most recently 
executed DTP (the last DTP on the active path), and the 
localized unknowns associated with it. It tries to find an- 
other way to expand a DTP along the active path that lets 
the listener know a localized unknown. The Analyzer backs 
up the active path testing each DTP in turn. If this fails, 
in the second phase, the Analyzer considers augmenting 
the existing DTP at the informational level. This level is 
reflected in the CTPs. The Analyzer examines the most fo- 
cussed DTP on the active path to see if it can he expanded 
with a CTP to let the user know a localized unknown. A 
detailed account of this type of processing is presented in 
\[5\]. 
184 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
co/DO (listener, =adopt ('Maple-plan)) 
 / com end (listener. DO(listerle~ ,*Maple-plan)) 
pt.A.~ ' ~ motive (listener, DO( listener, Msple-plaa)l 
SdZisC ( ..... ) / 
/ / ' -say( ACT-PLAN (avoid{l~avy-tramc), *Maple.plan))) 
Pl.;t./ V~h~rllslanllat~ ''taking I-tapl .... iris heavy traffic.'' 
t * "ACT-P~( at, old(heavy-frame) *Maple-plan))) say ( DO( listener. Maple-planD I 
'take Maple. ' ' ~.~I 
~/say( FEWER-BUSINESSES( *Maple-plan, *Sheridan-p\]anD 
' "since taking Maple there are fewer businesses than taking Sheridan. ' ' 
Figure 3: The TP for Before the Digressive Question in 
Demonstrations 2 and 3 
IS-DIGRESSION(f b) 
Let topic := IS-DIGRESSION-TYPE-I(fb) 
or 
IS-DIGRESSION-TYPE-2(fb) 
or 
IS-DIGRESSION-TYPE-3(fb) , 
If topic 
Then Let ctps =: get-ctps-with-nucleus(topic) 
Let related-reasoning =: get-reasoning(2,topic) 
Loop-for-each ctp in ctps 
If list ener-knows(effect-of(ctp)) 
or 
not(member(e~ect-of(ctp), related-reasonlng)) 
Then Let ctps := remove-from(ctps, ctp) 
End-loop 
If more-t han-one(ctps) 
Then Let etp-to-expand =: get-simplest-plan(ctps) 
Else Let tip-to-expand =: ctps 
If ctp-expansion 
Then post-content-goal(KNOW(listener, 
effect-of(ctp-to-expand))) 
ret u rn( ct p--to-expand) 
Else return('nU) 
5.4 An Example TP 
Figure 3 shows IDP's TP after its third response in Demon- 
strations 2 and 3. The TP has been formulated to achieve 
the DG of having the listener adopt the plan to take Maple 
Road. The high-level plan is a DTP which can decompose 
into other DTPs and CTPs. The TP always bottoms out 
in the primitive act, say. The argument to the say act is 
a text message which includes a proposition as the content 
to be expressed. 
In the TP, the checks (x/) mark the active path. Note 
that the plan for motivating the listener has not been ex- 
ecuted in the order indicated by the sequencing act snse- 
quence (see Figure 2(a))i In particular, to make the initial 
recommendation the planner selected a simple plan (advise, 
indicated by a dashed line) over a more complex plan (mo- 
tivate). In response to the listener's first why-question, IDP 
replanned the recommendation with the more complex mo- 
tivate act. Since its recorded TP indicated that it had just 
used the advise act, the only act in the plan for the motivate 
act that must be executed is the say act. This act conveys 
the satellite proposition ~that counts as the motivation for 
the advice. 
The second act in the plan for motivate, circumstantiate 
expands to an optional CTP. IDP does not expand and use 
this plan until it responds to Why? a second time. This hap- 
pens because IDP replans and expands its TP in reaction 
to the listener's questions. 
6 Detecting Digressive Questions 
If the Analyzer fails to recognize a way to replan Or ex- 
pand its own TP from the listener's feedback (Section 5.3), 
it tests the feedback to See if is a digressive question. Fig- 
ure 4 gives the algorithm that the Analyzer uses for the 
IS-DIGRESSION test. :The Analyzer tests the feedback 
(fb) in two stages. First, the Analyzer tests to see if the 
feedback has the correct form and content for one of three 
kinds of digressive questions that the system can detect. If 
one of the three tests succeeds, it will return a proposition 
that is the topic (topic) for the digression. The topic is 
the proposition that the listener questions in her digressive 
question. 
In the second stage, if there is a topic, the Analyzer col- 
lects CTPs (ctps) that use the topic as a nuclear proposi- 
Figure 4: Detecting Digressive Questions from the Lis- 
tener's Feedback 
IS-DIGRESSION-TYPE-1 (fb) 
If fb = "Why" {"not" } proposition 
and 
DONE(system, say(proposition)) 
Then return(proposition) 
Else return('nil) 
Figure 5: Test 1: For Detecting Digressive Questions Based 
on What the System Said Previously 
tion. It also collects domain reasoning that the system has 
performed that is related to the topic (related-reasoning). 
The Analyzer tests each CTP and removes any CTPs that 
have the effect of letting the listener know a proposition 
that she knows already. It also removes a CTP if its effect 
is not one of the related reasoning propositions. If there is 
still more than one candidate CTP, the Analyzer picks the 
simplest (ctp-to-expand). If a CTP has been found that 
meets all of the requirements, the Analyzer posts its effect 
as a CG and returns it for the Planner-Actor to expand 
to answer the question. If there is no such CTP, then the 
digression test fails. 
6.1 Detecting Digressive Questions About 
What has been Said 
The simplest kind of digression that IDP can detect is when 
the listener asks a direct question about a proposition that 
was expressed as part of the system's TP. The Analyzer uses 
the first digression test, IS-DIGRESSION-TYPE-I, that is 
described in Figure 5. If the feedback (fb) is a question in 
the form "Why" {"not"} proposition and the system has 
said the proposition as part of its TP, then the proposition 
is returned as the topic for the digression. Otherwise, the 
first test fails. 
The first test is used to detect the digression in Demon- 
stration i. Figure 6 gives IDP's TP just before the listener 
asks the digressive question in that example. When the lis- 
tener asks Why is there heavy traffic now?, the Ana- 
lyzer tries to find a way to expand the DTP-portion of its 
185 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
TP on the active path (Section 5.3). This involves searching 
the growth points in the DTPs for motivate, concede, and 
recommend. When this fails, the Analyzer tries to process 
the feedback as a digressive question. 
Testing for a digression (Figure 4) with the first digression 
rule (Figure 5), the Analyzer checks the TP in Figure 6 to 
see if the proposition in the listener's question was said 
by the system. The Analyzer finds that it has been said 
(Figure 6 in boldface). Therefore, the first digression test 
succeeds, and the following proposition is returned as the 
topic of the digression: 
OBJ-PROP(now, heavy-traffic) 
The Analyzer finds another CTP that uses the topic as a 
nucleus: 
circumstantiate(OBJ-PROP(now, heavy-traffic)) 
and collects reasoning that is related to the topic 
ACT-PLAN(avoid(heavy-traffic), *Maple-plan) 
OBJ-PROP(now, rush-hour)) 
SECO N DARY-GOAL-ACT(avoid(heavy-traffic)) 
The CTP's effect, 
KNOW(listener, 
OBJ-PROP(now, rush hour)) 
lets the listener know a new satellite proposition: now is 
rush hour. Since (he listener model does not assert that the 
listener knows this proposition, and since this proposition is 
related by domain reasoning to the topic, the CTP remains 
as a candidate for expansion. So that IDP's Planner-Actor 
can instantiate a plan for this CTP, its effect is made a CG. 
Then the CTP is returned for the Planner-Actor to expand 
into the TP. This leads to the restatement found in the last 
line of IDP's response in Demonstration 1. 
6.2 Detecting Digressive Questions About 
Inferred Propositions 
In Demonstration 2, IDP never expressed the proposition 
that the listener questions. Therefore, the listener must 
have inferred it. In this case, the system cannot determine 
that the question is digressive simply because the answer 
is something that the listener does not know. Such a rule 
would not distinguish digressive questions from incoherent 
ones. 
To test for digressive questions about propositions that 
the listener has inferred, the Analyzer uses the second 
test, IS-DIGRESSION-TYPE-2, that is described in Fig- 
ure 7. The Analyzer checks to see if the feedback is a why- 
question about a proposition (proposition) that has not 
been expressed by the system, but that can be inferred (IS- 
INFERABLE) from what has been said. If the proposition 
satisfies these requirements, then, as in the previous case, 
the proposition is returned as the topic for the digression. 
Otherwise, the second test fails. 
A question is coherent only if the listener has inferred 
it correctly from something that the system has said. Fig- 
ure 8 describes the test, IS-INFERABLE, that the Analyzer 
uses to determine if the questioned proposition is inferable. 
First, the Analyzer gathers all the propositions that IDP 
has said as the content of a say act (expressed-props). 
For each expressed proposition (prop), the Analyzer col- 
lects propositions that are within two reasoning steps of 
IS-DIGRESSION-TYPE-2(fb) 
fb = "Why" {"not"} proposition 
and 
not(DONE(system, say(proposition))) 
and 
IS-INFERABLE(proposition) 
Then return(proposition) 
Else return('nil) 
Figure 7: Test 2: For Detecting Digressive Questions that 
the Listener has Inferred 
IS- INFEI~AB LE(proposition) 
Let expressed-props := get-said-props 
Loop-for-each prop in expressed-props 
Let inferable-props:= 
union(get-reasoning(2, expressed-props), inferable-pr0ps) 
Endlo0p 
Let inferable-props := 
remove(expressed-props, inferable-props) 
If member(propositlon, inferable-props) 
Then return('true) 
Else return('nil) 
Figure 8: Determining if a Proposition is Inferable 
each expressed proposition. These are added to the propo- 
sitions that are inferable (inferable-props). Some of the 
propositions that can be inferred may have been said al- 
ready. Therefore, after collecting the inferable propositions, 
the Analyzer removes any proposition that is also one of the 
expressed propositions. FinMly, the Analyzer checks to see 
if the proposition that the listener has questioned is among 
those that are unsaid, but inferable, from what IDP has 
said previously. If it is, the test succeeds. 
In Demonstration 2, the second digression test is used 
to help determine that the listener's question is digres- 
sive. Figure 3 gives the system's TP just before the di- 
gressive question Why is there heavy traffic now?. In 
this interaction, the questioned proposition has not been 
expressed by the system. Therefore, the first digression 
test (Figure 5) fails immediately. In the second digression 
test (Figure 7), the form of question is satisfactory and 
the system has not expressed the questioned proposition. 
Therefore, IS-INFERABLE is invoked to see if the propo- 
sition 
OBJ-PROP(now, heavy-traffic) 
can be inferred from what has been said. 
At this point in the discourse, the propositions that IDP 
has expressed with say acts are 
DO(listener, *Maple-plan) 
ACT-PLAN(avoid(heavy traffic), *Maple-plan) 
FEWER-BUSINESSES(*Maple-plan, *Sheridan-plan) 
The Analyzer collects unsaid propositions that are related 
by reasoning to them. 
OBJECT-PROP(now, heavy-traffic) 
OBJECT-PROP(now, rush-hour) 
SECONDAR¥-GOAL-ACT(avoid(heavy-traffic)) 
N EWER-ROUTE(*Maple-plan, *Sheridan-plan) 
186 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
~/7om~nend(listener, DO(listener. *Maple-plan)) 
4 :on c~de(listener. DO(listener. *Sheridan-plan)) 
, ~ ..... _/,, v/mouv~te(listener, DO(listener, *Maple-plan)) condifiona"llze(D(~st~ner, *Sheridan~Uan)) -- | 
PLAN PLAN l 
• 'you could take Sheridan howe3;eT~ "' _ ~'. .. / - -~.j.5:~.-~ -, circu~0s~amiate(ACT/PLAN(avoid(heavy.ttaffic), 
FEWER INESSES Sheridan lall ~t y( ~,?USINE5 ( p -p . "-p )) 
• ' 'since ta ing Maple there are 
~'\[ ) fewer ~si?n%sMs%~ then taking / 
SheridCn/' 
• cir~zumst~fiate(DO(~stener, *MapleCplan)) / __ 
\[ \] 4~y(ACT-PLAN(avoid(heavy-traffie), *Maple-plan))) ~'~ 
/ •'taking Maple avoids heavy traffic.'' 
say(OBJ-PROP(now,heavy-t raffi~)) 
''now there is heavy cr~ffi_c.'' 
say(DO(listener, *Maple-plan)) 
: ' 'you should ~ake Maple. ' ' 
- Figure 6: The System's TP Just Before the Digressive Question - Demonstration 1 
IS-DIGRES SION-TYPE-3(fb) 
If ¢b = "Why"i {"not"} Then Let proposition := get-last-proposifion-s~id 
Figure 9: Test 3: For Detecting Garden-path Digressions 
These are the inferable propositions. Since the questioned 
proposition is among them, the test for the second type of 
digressive question succeeds. 
Next, the Analyzer tries to find a CTP that uses the 
proposition as a nucleus (Figure 4). As before, the CTP 
that it finds is 
circumstantiate(OBJ-PROP(now, heavy-traffic)) 
Since the listener does not know its effect, the Analyzer 
selects the CTP to answer the digressive question by making 
its effect a CG, and returning the CTP to be expanded and 
executed by the Planner:Actor. 
6.3 Detecting Garden-path Digressions 
Demonstration 3 shows that when the listener repeatedly 
asks Why?, IDP eventually recognizes a garden-path digres- 
sion. A garden-path digression will always occur at some 
point in an interaction, if the listener asks a series of simple 
why-questions. The test is a simple one that is given in 
Figure 9. If the feedback is Why? or Why not?, then the 
test succeeds and the last proposition that was expressed is 
returned as the topic of the digression. 
In Demonstration 3, When the listener asks Why? a third 
time, the state of IDP's TP is the same as in Demonstration 
2 (Figure ??). The first two digression tests fail because the 
form of the question is incorrect. Using the third digression 
test (Figure 9), the last proposition that the system ex- 
pressed, and therefore, the topic of the digression becomes 
FEWER-BUSINESSES(*Maple-plan, *Sheridan-plan) 
IDP finds a CTP that uses this topic as a nucleus 
circumstantiate(FEWER-BUSINESSES(*Maple-plan, 
*Sheridan-plan)) 
Reasoning that is related to the topic is 
ACT-PLAN(avoid(heavy-traffic), *Maple-plan) 
NEWER-ROUTE(*Maple-plan, *Sheridan-plan) 
The effect of the CTP is to let the listener know the sec- 
ond proposition above. Since the listener does not know 
this proposition, and since it is reasoning that is related to 
the topic, the CTP remains as a candidate for answering 
the digressive question (Figure 4). After asserting that the 
effect of the CTP 
KNOW(listener, 
NEWER-ROUTE(*Maple-plan, *Sheridan-plan)) 
is a CG, the Analyzer returns it to be expanded to answer 
the digressive question. In Demonstration 3, this results in 
the system response: 
since taking Maple is a newer route than 
taking Sheridan. 
6.4 Incoherent Questions 
In the IDP model, an incoherent question is a question 
that cannot be related to the system's TP either by replan- 
ning/expanding the system's TP, or by identifying it as a 
digression. Demonstration 2a shows how IDP processes in- 
coherent questions. When the Analyzer tries to replan or 
expand its DTP to answer the question (Section 5.3), it 
187 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
cannot find a TP that uses the propositional content of the 
question 
OBJ-PROP(now, heavy-traffic) 
as a nucleus. Therefore, analyzing the question to recognize 
a way to extend IDP's TP fails. 
Next, the Analyzer attempts to process the question 
as a digression (Section 6). IDP answered this question 
in Demonstrations 1 and 2. In Demonstration 2a, the 
questioned proposition has not been expressed previously. 
Therefore, the first digression test fails immediately. The 
third digression test also fails because the question is not a 
simple why-question. In the second digression test, the ut- 
terance satisfies the first two conjuncts. That is, it has the 
correct form, and it has not been expressed by the system. 
However, it is not inferable from what has been said so far. 
At the point in the discourse where the question is asked, 
the only proposition that IDP has expressed is 
DO(listener, *Maple-plan) 
Since this proposition is not a true fact in the knowledge 
base, no propositions are inferable from it. Therefore, the 
Analyzer determines that the questioned proposition is not 
a digression. Having failed all the tests, IDP responds with 
Huh?. 
Note that in Demonstration 2a, the proposition that the 
listener questions 
OBJ-PROP(now, heavy-traffic) 
and the answer to it 
OBJ-PROP(now, rush-hour) 
are among the localized unknowns (Section 5.1) that can be 
associated with the TP at the point in the discourse where 
the question is asked. This means that the information 
that the listener refers to with her question is highly rele- 
vant domain information. In spite of this, the question is 
incoherent. This suggests that discourse focus and purpose 
are more important factors for determining the coherency 
of a question than the domain structure. 
7 The System Implementation 
In the knowledge base, the several sources of information 
that the system needs to analyze and plan the discourse 
are all represented uniformly using the Semantic Network 
Processing System (SNePS) \[17\]. This includes knowledge 
of the text plan operators, the domain plans, entities in the 
domain, the user model, the discourse plan executed so far, 
and rules for reasoning about all of the above. The SNePS 
Actor models a cognitive agent operating in a single-agent 
world. It integrates inference and acting by representing 
beliefs, plans, and acts as structured intensional entities 
in the network formalism. Because the agent's world and 
planning knowledge is represented uniformly, the agent can 
discuss, reason about, formulate, and also execute its plans. 
Based on the TOUR model \[8\], the various driving routes 
that my system can discuss are represented as precon- 
structed plans that are composed of two types of actions: 
going and turning. The domain plans are represented at 
various levels of detail and, as conceptual entities, can have 
properties. Whenever the system reasons about the do- 
main, the reasoning that leads to deductions is recorded in 
the knowledge base along with the deductions themselves 
and is available as content for explanatiol,s. 

References 
\[1\] s. Carberry. A pragmatics-based approach to ellipsis reso- 
lution. Computational Linguistics, 15(4), 1989. 
\[2\] A. Cawsey. Generating explanatory discouse. In R. Dale, 
C. Mellish, and M. Zock, editors, Current Research in Nat- 
ural Language Generation. Academic Press, 1990. 
\[3\] H. P. Grice. Logic and conversation. In P. Cole and J. L. 
Morgan, editors, Syntax and Semantics 3: Speech Acts. 
Academic Press, New York, 1975. 
\[4\] B. J. Grosz and C. L. Sidner. Attention, intentions, and 
the structure of discourse. Computational Linquistics, 12, 
1986. 
\[5\] S. M. Hailer. Interactive generation of plan justifications. In 
Proceedings of the Fourth European Workshop on Natural 
Language Generation, 1993. 
\[6\] S. M. Hailer. A model for cooperative interactive plan ex- 
planation. In Proceedings of the The Tenth IEEE Confer- 
ence on Artificial Intelligence for Applications, San Anto- 
nio, Texas, 1994. 
\[7\] E. H. Hovy. Unresolved issues in paragraph planning. In 
R. Dale, C. Mellish, and M. Zock, editors, Current Research 
in Natural Language Generation. Academic Press, 1990. 
\[8\] B. Kuipers. Modeling spatial knowledge. Cognitive Science, 
2, 1978. 
\[9\] W. C. Mann and S. A. Thompson. Rhetorical structure 
theory: Towards a functional theory of text organization. 
TEXT, 8(3), 1988. 
\[10\] J. P. Martins and S. C. Shapiro. A model for belief revision. 
Artificial Intelligence, 35(1), 1988. 
\[11\] J. Moore and W. Swartout. A reactive approach to ex- 
planation. In International Joint Conference on Artificial 
Intelligence, 1989. 
\[12\] J. Moore and W. Swartout. A reactive approach to ex- 
planation: Taking the user's feedback into account. In 
C. Paris, W. Swartout, and W. Mann, editors, Natural 
Language Generation in Artificial Intelligence and Compu- 
tational Linguistics. Kluwer Academic Publishers, 1991. 
\[13\] J. D. Moore and C. L. Paris. Planning text for advisory di- 
alogues: Capturing intentional and rhetorical information. 
Computational Linguistics, 19(4), 1993. 
\[14\] J. D. Moore and M. E. Pollack. A problem for RST: The 
need for multi-level discourse analysis. Computational Lin- 
guistics, 18(4), 1992. discussion. 
\[15\] S. C. Shapiro and The SNePS Implementation Group. 
SNePS-2.1 User's Manual. Department of Computer Sci- 
ence, SUNY at Buffalo, 1992. 
\[16\] S. C. Shapiro, D. Kumar, and S. Ali. A propositional net- 
work approach to plans and plan recognition. In Proceedings 
of the 1988 Workshop on Plan Recognition, Los Altos, CA, 
1989. Morgan Kaufmann. 
\[17\] S. C. Shapiro and W. J. Rapaport. SNePS considered as a 
fully intensional propositional semantic network. In N. Cer- 
cone and G. McCalla, editors, The Knowledge Frontier. 
Springer-Verlag, New York, 1987. 
\[18\] J. van Kuppevelt. About a uniform conception of S- and D- 
topics. In Proceedings of the Prague Conference on Func- 
tional Approaches to Language Description, Prague, 1992. 
