Modeling Informal Debates

1. Conversational aoves

In the discourse theory proposed here, a conversation is seen as a sequence of lonyersatiQ~l ~. wherein each move corresponds o a spea er's ~unicatiye &g&l vis-A-vis a iarticular _prece ng sectjon of discourse. Support, · "Challenge, · and Interrupt, · are some or the communicative goals identified in this work, The goals are specific to ·maxim-abiding· [2J thematic development of a discourse, and <fo not refer to underlying speaker emotive intentions.

All conversational moves have a set of ( 1) precoo4itions which specifies the requisite

· This research waa supported in part by the Office of Naval Research under Contract No. N0001477-C-0371.

I'd like to thank Bill Woods, Lyn Bates, and Brad Goodman for their comments.

Rachel Reichllan

Harvard University and Bolt Beranek and Newman Inc.

discourse environment for the move's appropriate performance; (2) different .IIIRl1u. of fulfillment; and (3) varied effects on ~scourse context. Effective communication is enabled by conversants being aware of such a standardization of conversational moves.

2. A relevant discourse context

Discourse structuring and segmentation provides a mechanism by wnich to delineate that limited portion of preceding discourse directly relevant to the ongoing exchange. For example, when development of a context s~ce is interrupted, the interrupted space is not of direct relevance to the ensuing conversation until the digression is completed and the interrupted space is resumed from its point of interruption.

Different forms of topic suspension and resumption is a typical feature of extended discourse. By associating different levels of influential ~ with a context space, a space's various susp"iiiiiOn states during the course of an exchange can be captured.

The influential status of a space reflects its foreground-back&round discourse role. Only context spaces in a foreground role are directly related to current discussion, and only they are considered to be in the current relevant discourse context. In general, the relevant context ia lillited to the context space under current development (called the .a.c.t.1Y..e space) and the one other space (called the contro1HDf space) vis-A-vis which the ~c~o~mm~un~l~c~a~t~ve goal of the active space applies.

In the context space framework, relevant context identification is enabled by recording the standardized effects of taken conversational moves. These effects include reassigning the influential status of preceding spaces and creating and deleting outstanding discourse expectations of moves to follow.

3. Context spaces

Context spaces encode information both explicit and implicit in the dialogue. The spaces are characterized in much the same way as elements of a ·systemic Grammar· [4] via attributes represented as ·slots· A la Minsky [8]. Some slots are standard for all context spaces. Other slots are specific to the communicative goal served by the space - they reflect the implicit inferential elaborations needed for a constituent type to serve a given discourse role.

For example, substantive supports usually involve expression of some state-of-affairs, F, such that there exists some generic support­ erinciple, P (A la Toul11in's notion of a backing· [ 15] )l of the form "A Implies B, · such that F is an nstance of A (or Not B), C is an instance of B (or Not A), and using some inference rule like M'ildus-Ponens (or Modus-Tollens), C follows from F ·

In illustration~~.. consider R' s utterances on Lines 1-16 below. 1·he excerpt is taken from an informal debate between friends, where the "nature­ nurture" controversy is under current contention. R's utterances are in response to D's preceding claim that most criminal behavior is not genetically determined.

Excerpt l.

R: 1. 2. Except however, John and I just saw this two hour TV show,

where they showed - it was an excellent

H: R: ~: Uh hum

the child is not

5. 6. French TV documentary - and they showed that, in fact, the aggressive nature of

b: influenced by his environment.

really that much

M: 9. How did they show that?

R: 10. They showed that by filming kids in

11. M: 12. Uh hum,

kindergarten,

R: 13. M: 14. showing his behavior among other children, And then?

R: 15. 16. and showed him ten years later acting the

D: 17. 18. ~:f!,w~~·c~~~~?sth~~ls where he learns his oehavior, in kindergarten.

H: 19. Oh, sure.

R: 20.

21.

Oh, and it was twins. The important thing

was that there were two children from the

same environment, whereas only one of the

22. 23. brothers acted tnat way.

211. 25. So you couldn't blame it on the child's home.

D: 26.

27. R: 28. It has to do with the child's environment. Right, but the two brothers have the same environment.

D:

R: 1.

D: 2. ~~: Because you and I are ver~ close in this room right now but we don t have the same environment.

It has nothing to do with the child's home.

2 ~6: They do not ·have the same environment. Why not?

~~: Because I'm looking at you, I'm seeing that window behind you, you're no.t seeing

that window behind you. You are not

g, o. Two people canlt be in exactly the same place at the same time, otherwise, they'd

41. 42. occupy the same space. They do not have the same environment.

~ b: looking at you I am doing it.

43. They don't have the same friends.

46. 47. family is really aeaningless 1 because when you think of the difference n treatment

H: 411. 45. And, I mean, they don't even - You know, to say that two kids come from the same

48. that two kids can get in exactly the same

49. family it's incredible. You know, it's

50. the difference between night and day.

On Lines 1 -16, R performs two conversational moves: ( 1) the assertion of a claim (Lines 5-6) ; and (2) citation of a support for this claim. The support move involves both an authority citation (Lines 1-5) and a substantive sup_port analyzed as follows (using the abbreviations~ C (Claim), F(Fact of Supporq, 1' (Principle), M ~Mappings between P and C), M (Mappings between P and F), and LR (Logic Rule)):

• c = Not(The aggressive nature of the child is influenced by his environment) - Not(A)

· For other systems using rules formalized semantic relations to connection between utterances argumentation and ordinary see ( 1, 13, 5]. of logic and explicate the in informal discourse

• F = Not(A child's aggressive behavior in kindergarten changed over ten years) - Not(B)
• p = IF (A) a person's social interactive behavior is environmentally influenced Then (B) that person's social interactive behavior will change over time
• M = ((person's social interactive child's aggressive behavior)) behavior,
• H' = ((person's social interactive behavior, child's aggressive behavior), over time, ten year duration period from kindergarten))

LR = Hodus-Tollens

All supportive context spaces are characterized by slots to hold a Source, Method, Credentials, Access, F, P, M, and M' components of a support move.

4. Surface linguistic phenomena

Effective conversational engagement rests on listeners being able to identify the functional relation of a speaker's utterances to the Preceding discourse. A speaker 1 s surface linguistic forms often signal the relation involved.

~~:

Speakers often preface their substantive l,ltteraz;~ces with linguistic connectives whose sole (major} import seems to be conversational move marl<ing ( 7, 11 , 14]. For example when directly attacking an opponent, a speaker wih say something like, "No, but," whereas, when indirectlt attacking, the speaker will say "Right/Yes, but (e.g., Line 29, Excerpt 1). The "right/yes" of the indirect challenge indicates that while the opponent's claim may be true, it is basically irrelevant or overly weak to be of major import to the argument. The "no" of the direct challenge explicitly undermines the very validity of tfie opponent's claim.

(Non)Prongpioalization:

Traditional theories of pronominalization based on "recency" and "semantic ambiguity" criteria cannot account for the many cases of nonpronominalization that one finds in spontaneous disc9urse. An effective model of discourse (non}pronominalization must be based on recognizing a discourse's underlying structural, thematic organization, and the interrelationship between this structure and the discourse "focus". The predominant pronominalization rule encoded in the context space theory is that only elements in high focus in a current relevant discourse context will be pronominalized ,

The nonpronominalization phenomenon highlighted here, I believe, exhibits the. high level of subconscious rule-governed Processing involved in discourse engagement. For example, in all probability, most readers, consciously analyzing Excerpt 1 will assume that the nonpronominal cases clted could all be replaced with pronominals without any so called grammatical violation or obscurity of intented referent. Yes, the intended referent would probably be retrievable without undo difficulty. However, in the course of the exchange, such pronominals would convey subtle conflicting messages to a listener about the speaker's underlying structural · representation of the discourse. In addition these cases are but prototypical of the repeateJ nonpronominalization found in discourse. The phenomena warrants explanation.

For example, consider the contrasting cases of (non) pronominalization on Lines 24, 29, and 43 of Excerpt 1.

.I.J.J:ul. .z!i.: Line 24 is supported by Lines 20-23

and is used by R to modify and supersede her initial claim of Lines 6-8. After acceptance of D's Lines 17-18 challenge, only a child's environment before kindergarten is relevant. R therefore eauates "environment before kindergarten" with "child's home" and addresses herself to this residual issue.

The antecedent of "it" on Line 24 is "the aggres~ive nature of the child," last mentioned on Line b. A relevant discourse context for a replacement claim is the claim being superseded. Using such a structural analysis of the discourse enables effective modelling of R's long distance pronominalization.

1..1ne. .z.ll.: This same criteria of discourse pronomrnBlization and structural analysis of a discourse enables effective modelling of R's subsequent short distance nonpronominalization of "the two brothers" on Line 28 though they were just referenced on Line 21 and no other potential antecedents had intervened. By Line 28, Line 21 is no longer part of the relevant discourse context.

• o Concluding a set of supportive utterances (Lines 20-23) with the claim they support (Line 24) removes the supportive utterances from the relevant discourse context.

~ 42: Line 42 is a restatement of Line 30 and Line li3 is a further support of it. (The utterances cannot be a continuation of Lines 39-41 because they are of definite modalityt whereas the modality of Lines 39-41 is hypothetica!.)

Despite D's intervening use of "they" to refer to "two people," his use of "they" on Line 42 to refer to "the two brothers" is unambiguous, because the utterances containing "two people" and the preceding "they" are no longer in the current relevant discourse context. They are therefore not possible antecedents for this pronominal form.

• 0 Resumption of the initiating subject of an analogy (Line 42) simultaneously reinstantiates the initiating utterances as the current relevant discourse context and removes all analogous and transitional utterances from it (Lines 32-41) .

c. The ATN Grammar

The formalism chosen to encode the context space process model is an Augmented Transition Network (ATN) A la Woods [ 16] register testing and action executions along ATN transitions seem a natural way capture many of the context sensitive aspects of extended discourse. In the discourse ATN, these tests and actions correspond to the preconditions and effects of conversational moves, and state transitions, often, correspond to a finer characterizttion of conversational move mode fulfillment ·

Figure in Appendix I represents a small portion of the discourse ATN. (Due to space considerations only "Push"/"Jump" actions are noted.) The model is written in terms of a basic loop, a single traversal of which represents the taking of a single conversational move. On each transition cycle, the process model begins in the "Produce-Next-Hove" state, wherein it chooses a conversational move category; in one of the last arcs traversed, there is aelineation of an abstract characterization of the message to be generated. The characterization specifies a semantic/logical relation between the utterances generated and utterances contained in the current controlling context space. This characterization must be met if the subsequent utterances are to fulfill the given conversational role category. the characterizations are written in a predicate­ calculus like lan1uage, where functions like "For All, 11 "Infer," Imply .I" and "Exclusive-OR," are considered primitives o~ the system.

· See [ 12] for full discussion of further correspondences, and distinctions, between this discourse ATN and typical ATNs used in sentential processing.

There are three major modes of processing performed in a conversation: ( 1) finer and finer categorization of the type of message to be generated; (2) updating our mental models of the discourse in preparation for generation of such remarks, and (3) actual production, Correspondingly, the grammar's states (via the trpe of actions on tneir arcs) are of three modes : ( 1) organizational, wherein tests and decisions are performed in finer categorization of the forthcoming conversational move category; (2) oonatruotive, wherein the updating actions associated with a given move, which include register assignments and the creating and updatinf of context spaces, are performed; and ( 3 productive, wherein the actions on arcs ental formation of the message to be generated.

1. Modelling informal debate moves

In informal debating, our highest level categorization of likely conversational moves are (1) challenging an opponent (executing a challenge­ choice expectation); (2) further challenging an opponent (executing a further-challen~e expectation); and {3) giving support to oneTs claims. The following are tne initial discourse model updating effects individually associated with each move.

The Challenge Hove: A next speaker is selected who is either: ( 1) the speaker noted in the accompanying discourse expectation; (2) a conversJ~.nt known to be on the expected speaker's side; (3) a previously neutral conversant who is appended onto the expected speaker's side; or (4) a conversant switching sides in the debate. In addition, the conversant being challenged (i.e., the last speaker) is assigned the role of Future­ Defender1 to reflect that s/he will most likely countercnallenge the forthcoming challenge.

The Further-Challenge Hove: A furtherchallenge always entails a discourse pop to some preceding claim of an opponent which has led to current subargumentation. It therefore involves closing intervening argumentation and re­ establishing the context srace po,Pped back to (or one of its direct subconst tuents) as the relevant discourse context for subsequent generation. This is accomplished by traversing tne links between context spaces, beginning with the active space and ending with the one being popped back to (i.e., the context space {loted in the discourse expectation being processed),

A next speaker is then chosen in manner similar to the counterchallenge case. Here however, the last speaker (who is a team member of the newly selected speaker) is not assigned the Future-Defender role, but rather, the Protagonist of the space popped back to is.

The Support Choice Hove: Here too the grammar's first action is to select a next speaker. In this case~ either the current speaker continues to hold the Iloor, or a speaker known to be on the current speaker's side is selected as next speaker. As a new challenge is not being generated, no conversant is put fnto the Future-Defender role.

2. Tracing Through Excerpt 1

To illustrate the operation of the Discourse ATN, let us follow some details of the process · modeling Excerpt 1.

R's move on Line 1 of the excerpt, as indicated by the clue words, "Except, however,· is a challenge conversational move. At this point in the conversation the conversants are already in the midst of a debate, and R is counterchallenging D's preceding claim that most criminal behavior is not environmentally induced. Let's begin then with the grammar having already decided to have R as next speaker performing a challenge conversational move. Major features of the grammar's current model of the discourse, then, are:

• o HEAD-CCS points to C1, the current controlling context space in the discourse environment.
• 0 Major slot fillers of Controlling; TYPE = Protagonist = R; Claim are inherently evil; CounterClaims = C2; C1 are: State = Debative-Issue; = Many criminals Antagonist = D;
• o CCS points to C2, the current active context space in the discourse.
• 0 Major slot fillers of C2 are: TYPE = Deoa ti ve-Issue · Protagonist = D; Claia = Genetics is no! to blame for most people's criminal actions· Goal = Challenge; Co­ Relator = C1; He!hod = Range-Application; State = Active.
• 0 SPEAKER = R· FUTURE-DEFENDER = D· PARTICIPANT-LIST = (D, M); SIDES ~ ((D)i(R)); EXPECTATION-LIST = ((Further­ Chal enge, D, C1)).

The process model begins by appending onto its list of outstanding discourse expectations the expectation that R is likely to later further challenge D's preceding claim. Then, modeling R's move, it chooses an Indirect-Challenge mode of argumentation.

Indirect challenges (as contrastive with some forms of direct challenges) always entail citation of a substantive claim whose communicative goal is in direct opposition to that of the current active context space. As such, this conversational move category always results in a context space shift and establishment of a new relevant discourse context. To set the proper environment for subsequent generation, therefore, the grammar now pushes to a constructive state, whose actions include:

• o changing the state value of the context space pointed to by HEAD-CCS (i.e., C1) to �enerall~· A Generating state reflects that co~inued discussion is indirectly related to the context space, and that therefore there is s~e chance of a later return to this space las reflected by the (Further-Challenge, D, C1) expectation yet outstanding);
• o changing the state value of the context space pointed to by CCS (i.e,, C2) to f~~trollipf, and assigning HEAD-CCS to it. s re ects that while subsequent generation will be directly related to the space, it will not be a continuation of its contained utterances.
• o creating a new debative-issue context space, C3, to contain components of the forthcoming move. CCS is assigned to it, and its following slots are filled: State <-: Active; Goal <- Challenge; Co-Relator <- C2; Protagonist <- R.

Returning from this constructive state, the grammar is now ready to further subcategorize the mode of indirect challenge. Since the modality of D's claim is negative, a challenge based on Hodus­ Tollendo-Ponens is appropriate li.e., in response to an opponent's claim of Not(B), a speaker claims Not(A), where A EXCLUSIVE-ORB is true) · .

The grammar begins by generating the clue words associated with this move, "Except however " and, then, facilitates citation of an aulhority for the claim about to be given. Supportive utterances and the claim they support serve different functional roles in a discourse and as such are distinct constituents of it: supportive spaces are subconstituents of the space they support. Therefore, in preparation for R's support citation the grammar first constructs a new supportiv~ context space, C~, assigning its Goal slot to Support and its Co-Relator slot to C3. R's authorily citation is then generated, and inserted into the authority slot of this space.

Returning from the support push the grammar fills C3 's SupportCS slot with the ~wly created support spacel and then it generates R's challenging clam. The claim is inserted into C3's claim slotk and "Modus-Tollendo-Ponens" is inserted into its ethod slot. Then, as common for all indirect challenges the new challenge context space is inserted info the CounterClaim slot of the space being challenged (i.e., the controlling context space C2).

The grammar then pops back to the general challenge state category, wherein it always creates the expectation that the · challenged conversant (i.e., the one assigned to a Future-Defender role)~ will counterchallenge the challenge just citea (i.e., the active context space).

Having ended simulation of R's challenge move, the grammar re-enters its start state where it begins to process the next conversational move. There it chooses to process this last discourse expectation created (i.e., (Challenge-Choice1, D, C3J). Control therefore once again immediately passes to the general challenge state of the grammar. This time, M, yet an unaligned conversant, is chosen ae next speaker and she is appended onto D's side of the debate. R the preceding speaker 1 18 assigned the Future-Defender role. 1:ontrol ~s then passed to the challenge directly state.

Since R's support of C3 only entailed citation of an authority, demanding some substantive support, via generation of a "how" or "why" question, is appropriate. As this move does not involve citation of a contrary claim, the move does not result in the creation of a new context space or accompanying establishment of a new relevant discourse context. Instead, its effect is to restrict the addressed conversant's subsequent conversational move to one of support . This constraint is noted in the counterchallenge expectation set up for the Future-Defender (i.e. ( Challenge-Choice1 1 R.l C3, "Supply-Support-Claim"~ is appended onto tne ~xpectation-Llst register) .

On the grammar's next cycle through the network, it chooses to process this expectation; R is selected as next speaker, and H is assigned the role of Future-Defender. This time, however transitions to either the direct or indirecl challenge forms are forestalled by the Expectation register test on the arcs of these transitions. In accordance with maxim-abiding constraints, the grammar, therefore, goes to the Support state wherein it generates the demanded support (a formal analysis of which has been cited earlier) .

At this point , let's skip to the last section of the excerpt in illustration of the grammar's method of modeling context space resumptions. H's utterances on Lines ~~-50 do not continue D's preceding conversational move, but rather, serve as a further challenge to R' s claim on Line 2~, via attack on its support of Lines 20-23. Formally speaking, H's challenge entails challenging R's mappings of "two people sharing the same environment" onto "two twi ns living in a same home."

As such, the relevant discourse context for subsequent generation should be C6, the supportive context space about to be challenged. However, currently, C6 1 s state is ~. and the relevant discourse context is compo~ D's debative-issue context space ( C1 0, which contains his clai}ll Qf Line 30/42) and its supportive context space (C1b, which contains Line 43 of the excerpt). To remedy this situation the grammar must perform some extensive discourse model updating.

H's conversational move on Line 44 is a further-challenge. As discussed, the first action of this move is to close all intervening subargumentation stemming from the space being popped back to (as noted in the discourse expectation) up to and including the current active space. This is accomplished by traversing the network links between these context spaces. l1ajor features of the grammar's context space configuration by Line 45 (of relevance here) are:

C6: Line 20-2~; Goal: Support; Method: Hodus­ TOllendo-Tollensl; Co-Relator: C6; State: Closed.

cr: Line 24; Goal: Fix-Claim; Method: ~nstrainment; Co-Relator: C3; SupportS: C7; CounterClaims: C8; State: Generating.

C8: Lines 26-27; Goal: Challenge; Method: Irrelevance-By-Expansion; Co-Relator: C7; CounterClaims: C9i State: Generating.

~: Lines 28-~9· Goal: Challenge; Method{1}: Apply-Expansion; Co-Relator{1}: C8· Method{2}: Derived-From; Co-Relator{ 2}: C7; C!ounterClaims: C10; State: Generating.

~: Lines 30 42; Goal: Challenge· Method: Deny­ TrUth; Co-Relator: C9; SupportS: (C~1, C16) State: Controlling.

.C.U: Lines 32-34; Goal: Support j Method: Analogy; "CO-llelator: C10; SupportS: (Cl2, C13); State: Closed. ·

CJ2: Lines 35-37 · Goal: Support; Method: Hodus­ YOrlendo-Tollens; Co-Relator: C11; State: Closed.

~: Line 38; Goal: Support; Method: HodusYOrlendo-Tollens; Co-Relator: C11; State: Closed

tl..!l: Lines 39-40; Goal: Relate-Analogy· Method: IDBtraction; Co-Relator: C10; Supportts: C15; State! Closed.

~: Line 41· Goal: Support· Method: ModusYOrlens~ Co-Relalor: C14; State: bosed.

CHi: Line 43; Goal: Support; Function: Modus­ YOrlens; Co-Relator: C10; State: Active

Beginning with the active context space, C16, the grammar starts closing subarguments by reassigning its state and that of context spaces C10, C9, and C8 to Closed. Notice that intervening discussion contained in context spaces C15, C14, C13, C12, and C11, are not accessed in this traversal. This correctly reflects the influence of closed context spaces: they are entirely irrelevant to subsequent discourse processing (unless they are explicitly re-entered).

Having popped the discourse, the grammar now can select M, already aligned on D's side of the debate, as next speaker, and assign R, the Protagonist of the popped back to space, to the role of Future-Defender. The grammar now has to further subcategorize the type of further challenge to be process~d. Two main subcategorizations are possible: (1) new counterclaim/countersupport specificati-on; or (2) further support of a preceding counterclaim/countersupport. M 1 on the offense, chooses the former: CCS is reset to C7 1 and processing continues as for initial challenges - state . transition to Challenge-Choice2, wherein the direct challenge form is chosen.

As H's challenge involves challenging specifics of R's preceding support for C7, control is passed to the support-specifics-challenge state, wherein, C6l C7's support space, is reinstantiated as control! ng. A new debative-issue space is then created to contain M's subsequent utterances. (For further details of the force of H' s argument, and the grammar's analysis of it, see (12].)

D. Conclusion

The key element of this analysis is a structural and functional decomposition of a discourse into distinct, but related and linked, context space constituents. Context spaces are characterized by their functional role in a discourse and they contain information both explicit and implicit in the dialogue. This decomposition is enabled by:

• o formalization of a set of standard conversational moves performed in discourse;
• o formalization of the effects of each of these moves on a discourse context. Such effects constrain and set up predictions for subsequent thematic development;
• o identification of clue word cues that signal a conversant's forthcoming conversational move;
• I have illustrated that such a structural decomposition enables the grammar to:
• 0 effectively model discourse (non) pronominalization;
• 0 identify the limited preceding section of "talk" serving as a point of reference for subsequent discourse generation and/or interpretation;
• 0 clearly mark and hence ignore all intervening discourse that need not even be considered in re-establishment of a relevant discourse context;
• 0 have explicit annotation of elements of a discourse that are only implicit in a speaker's verbalization of a given conversational move. Such explication enables the grammar to identify those components of a freceding move that an opponent is chal enging, though at a superficial level, ttie opponent <loes not deny anything that has been said.

The context space theory delineates a single abstract structure underlying all discourse forms - expository text, argumentative text, narrative text - and based on such structure characterization it is able to specitt a single set of "maxim-abiding," "well-formedness, rules applicable to, and governing all discourse forms. Its basic components of segmentationL selective attention, relevant context identii·ication, cues, and expectations, correspond to the fundamental elements of many cognitive processes (9, 10].