A Compuaional Model of Social Perlocuions David Pauler and Alex Quilici Universiy of Hawaii a Manoa Deparmen of Elecrical Engineering 2540 Dole S Holmes 483 Honolulu, HI 96822 Absrac The view ha communicaion is a form of acion serving a variey of specific funcions has had a remendous impac on he philosophy of language and on compuaional linguisics Ye, his mode of analysis has been applied o only a narrow range of exchanges (eg hose whose primary purpose is ransferring informaion or coordinaing asks) while exchanges mean o manage inerpersonal relaionships, mainain ufacen, or simply o convey hanks, sympahy, and so on have been largely ignored We presen a model of such Usocial perlocuions" ha inegraes previous work in naural language generaion, social psychology, and communicaion sudies This model has been implemened in a sysem ha generaes socially appropriae e-medl in response o user-specified communicaive goals 1 Inroducion The imporance of viewing uerances as no simply saemens of fac bu also as real acions (speech acs) wih consequences has long been well undersood (Searle, 1969; Ausin 1975; Grice 1975) As a resul, i is imporan o sudy no jus he formal aspecs of language forms bu also how speakers use differen forms o serve differen funcions For example, one funcion of he ac of informing anoher person is o make he person aware of a sae of affairs; similarly, one funcion of promising is o secure he reurn of a favor Unforunaely, he sudy of speech acs has been largely limied o he collecion and clmmificaion of ac ypes and he condiions for appropriae use of each ype (Searle 1969; Wierzbicks 1987) The range of funcions, or perlocu~io;~ry effeef~, served by differen ac ypes has been largely ignored In paricular, here has been lile or no work on he impac ha speech acs can have on social aiudes and behavior Ye, wihou an accoun of how commu- Professor WHITNEY, Thank you for your inviaion ] Unforunaely, I will no be able o give a alk a THE U OF M COMPUTER SCIENCE DEPARTMENT on APRIL 14, 1998 I regre ha I mus decline ] I have a previous commimen [ You may wan o invie DAN VLASIK in my place! He is well-acquained wih he work we do here ] a McCORMICK SYSTEMS If you would like o pursue his opion, please conac him direcly a (808) 555-1973 Figure 1: A LeerGen Oupu Sample nlcaion can affec social siuaions, i is impossible o consruc sysems ha are capable of generaing socially appropriae ex This paper provides a compuaional model of aocial perlocuionj, and i describes how his model has been used o consruc an auomaed sysem, eergen~ for generaing socially appropriae e-mall messages and leers This sysem akes general communicaive and social goals from he user, such as demanding acion or expressing congraulaions, queries he user abou subgoals and perinen background informaion, and generaes he ex of an appropriae message by planning individual speech acs As an example, Figure 1 shows a message generaed by LeerGen in response o an inpu goal o decline an inviaion poliely In his example, he wrier was invied by he addressee o ravel and give a alk, bu he wrier had a previous commimen and mus decline However, he wrier knows some- I I 1020
one who could give he alk in his place The sysem planned s se of speech acs and realized each as a clause or phrase using a ex emplae library These acs include (1) hanking, (2) declining-reques, (3) apologizing, (4) making-excuse, (5) advising, (6) assuring, and (7) requesing Mos of he ex in he leer is devoed o addressing he wrier's social goals of being polie and helpful In conras, a leer wrier concerned only wih informing he addresee ha he was no paricipaing would likely say lile oher han "I won' be giving a alk a your even n, a socially inappropriae response 2 Previous Research Our work builds on resuls from hree disparae areas: naural language generaion (NLG), communicaion sudies, and social psychology The NLG communiy has focused on a small subse of he five generally acceped caegories of speech acs (Levinson, 1983): 1 Represenaives--saemens given as rue depicions of he world (eg, assering, concluding) 2 Direcives---saemens aemping o persuede he hearer o do somehing (eg, ordering, advising, warning) 3 Commissives----saemens ha commi he speaker o a course of acion (eg, promising, acceping a reques, aking a side) 4 Expressives---saemens expressing a psychological sae (eg, apologizing, congraulaing, condoling) 5 Declaraions---saemens effecing an immediae change in he insiuional sae of affairs (eg, chrisening, firing from employmen) In paricular, research in NLG has been limied o one ype of represenaive (ie, informing) and one ype of direcive (ie, requesing), and i has furher focused on informing's poenial o con~/nee he hearer of some fac and requesing's poenial o persuade he hearer o do some acion (Allen e al, 1994; Appel, 1985; Bruce, 1975; Cohen and Perfaul, 1979; Hovy, 1988; Perraul and Allen, 1979) As a resul, i has largely ignored speech acs in oher caegories, such as promising, advising, and crediing, as well as heir poenial perlocuionary effecs of creaing airnniy beween speaker and hearer, securing fuure favors for he speaker, and so on 1021 In conras, research in communicaion sudies has explored sraegies for persuading, creaing affiniy, comforing, and many oher inerpersonal goak (Daly and Wiemann, 1994; Marcu, 1997) For example, he sraegies for persuading include no only requesing, bu also exchange, ingraiaion, and sancions However, hese effors have no analyzed hese sraegies in erms of speech ac ypes and perlocuionary effecs so ha hese sraegies migh be realieed in compuaional form Finally, research in social psychology has looked a how personaliy rais affec inerpersonal ineracion For example, Kiesler (1983) formulaed general rules for describing how one person expressing one rai (eg, merciful) can lead o anoher person expressing a symmeric and complemenary rai (eg, appreciaive) Such ineracion dyads are direcly msppable o he speaker/hearer dyad of speech ac heory, and he vocabulary of rai erms and predicive rules sugges one way of lending organizaion o he grea variey of perlocuionary effecs Ye, social psychologiss have no mapped heir general rai erms o he classes of speech acs ha migh express hese rais Wha's been lacking is an aemp o inegrae he lessons learned from hese differen research effors o provide an iniial model of social perlocuions; ha is, a model ha describes how specific speech ac ypes have he poenial o produce specific effecs in a hearer corresponding o a speaker's social goals, and ha is specified formally enough o be used as par of ex generaion sysems 3 Our Model There are wo key quesions o address in forming a compuaional model of social perlocuions: Wha are he possible socially-relevan effecs of speech acs? Wha are he relaionships beween differen effecs? 31 Social Perlocuionary Effecs We have developed a axonomy of social perlocuionary effecs of speech acs These effecs are defined in erms of menal aiudes of he hearer, following he assumpion in speech ac heory ha all perlocuionary effecs follow from he hearer's recogniion of he speaker's communicaive inen The axonomy is:
6 Beliefs abou speaker's precise communicaive conen and communicaive inen Beliefs abou he speaker's inen o benefi or harm he hearer Beliefs abou he heaer's or speaker's responsibiliies (ascribed or underaken) Beliefs abou (or, impressions of) he speaker's personaliy rais The heaer's emoions The relaionship beween he hearer and he speaker 7 The hearer's goals We developed his axonomy by reviewing he communicaions sudies and social psychology liersure, as we]] as by analysing a corpus of leers and e- nudl messages for heir speech acs and mos prominen social effecs Prior research on speech acs has largely ignored several of hese caegories, especially he effecs on personaliy impressions, emoions, and he speaker-hearer relaionship 32 Relaionship Beween Social Effecs This axonomy is imporan because here appear o be significan resricions on he relaionships beween hese differen classes of effecs Figure 2 shows how hese differen ypes of effecs are relaed The arrows represen poenial causal links beween effecs These links are poenial because here are specific condiions associaed wih specific effecs ha dicae wheher one effec will cause anoher Essenially, he effecs sar wih he hearer's recogniion and accepance of a message's conen and culminaes in changes o hearer goals and he relaiouship beween he hearer and he speaker Tha is, a speech ac direcly resuls in beliefs abou he conen and inen of uerances and hese beliefs indirecly resul in changes o goals, emoions, and inerpersonal relaionships Specficially, hese belief can lead o indirec changes in he heaer's belief abou he speaker's inen o benefi or harm he hearer, as well as changes o he heaer's responsibiliies ha involve he speaker In urn, changes in belief abou wheher he speaker inends o benefi or harm he header can lead o changes in he hearer's goals, he heaer's emoions, and he heaer's impressions of he speaker's personaliy rais Finally, 5 changes in H's goals 6 / changes in H's emoions 2 H's belief abou f S's inen o benefi or harm H H's belief in conen and inen of acs direced a H T S's speech acs direced a H 7 changes o he srengh of H's relaionship wih S 4changes in H's impressions ors'srais 3 changes o H's beliefs abou responsibiliies involving S Figure 2: The Relaionships Beween Social Effecs changes o he hearer's emoions can lead o changes in he hearer's relaionship wih he speaker Our hypohesis is ha Figure 2 provides a framework ino all speech acs wih social effecs can be mapped To es his hypohesis, we analyzed in deail he relaionship beween he effecs of 40 differen ypes of speech acs, and we successfully placed each ino his framework (Pauler, 1999) These speech acs were ypical of he leers and messages we colleced, and hey were represenaive of four of he five main caegories of speech acs1 Figure 3 is an example, showing hese effecs for apolo~zing, a Alhough no shown in Figure 3, he causal relaionships beween hese effecs have condiions aached o hem In Figure 3, for example, a condiion on an apology leading o he hearer believing he speaker feels regre is ha he hearer believes he speaker is sincere and here is an ac for which 1We did no represen deelar6ions because we chose o focus m acs used in casual, inerpersonal ineracions raher ha~ acs ha were insiuionally framed =We do no rl;,, ha he model applies o groups oher h~n adu/ Wesea'ne~ See Bm'nlund (1989) for comparisons en he use of differen speech acs by Americana and Japanese 1022
"0 ~ 0 H's relaionship wih S is srenghened H's liking for S increases fecs of social speech acs leads o he quesion: How can we efficienly generae he speech acs we need o achieve an appropriae emoional response in he hearer? 4 A Model Of Leer Generaion 0 S is likable Praising Denying S is conscienious l ~ praise S is accounable l ~o o o S feels regre Apologizing Figure 3: The Effecs Of Apologising Warning Thanking an apology is appropriae We draw our erminology for describing specific personaliy rais (eg, likeable, conscienious) and emoions (eg, graiude, liking) from exising axonomies (Kiesler, 1983; Orony e al, 1988) Figure 3 shows effecs wih arrows leading o hem from oher speech acs, such as praising, warnhag, hanking, and so on These speech acs are here o illusrae ha speech acs are relaed hrough a web of inerlocking effecs Tha is, he causal relaiouships beween speech acs and effecs is manyo-many: a single ac can have many differen effecs and any single effec can be brough abou by many differen acs For example, expressing a demand can bring abou compliance, anger, or boh, and similarly, anger can be caused by a variey of oher acs, such as issuing a hrea In Figure 3, boh praising and apologizing are examples of acs ha can increase he heaer's liking for he speaker, and boh apologimng and hanking can lead he hearer o believe he speaker is accounable This large web of relaionships beween he ef- To illusrae he power of our model of social perlocuion, we have applied i o he ask of e-mail generaion in a sysem called LeerGen The sysem's primary ask is o ake a high-level communicaive goal (eg, inform a colleague ha one can' aend a meeing) and sugges a se of speech acs o achieve ha goal However, once i has made his suggesion, he sysem hen inerac wih he user o deermine which speech acs will appear in he final message and o acquire any addiional bakground informaion needed o iusaniae senence ex emplaes associaed wih each speech ac In addiion o he user's explici inpu goal, he sysem works wih a se of "sandardn user goals These goals fall ino hree classes: 1 Cos avoidance avoiding undesired aspecs of a curren or incipien siuaion, such as unwaned social percepions of oneself 2 Saus-quo mainenance ~elecion of an ac because one of is effecs would reinforce a desired aspec of he curren siuaion (eg, offeing o help anoher person because i would reinforce one's self-image as a generous person) 3 Trai-based habi--performing of an ac as a imeworn expression of a personaliy rai These goals can be hough as a sereoypical model of he user (Chin, 1989) These goals are achieved opporunisically during he process of deermining speech acs for he explicily provided user goal 41 A Graph-Based Represenaion Of Speech Ac Relaionships LeerGen essenially represens he perlocuionary effecs of speech acs as a large graph Figure 4 illusraes a porion of his represenaion ha relaes he speech acs of declining, hanking, and apologizing The nodes of he graph represen various effecs, and he unlabled edges represen a causal relaionships beween wo effecs There are also consrains on when edges can be raversed (alhough hey are 1023
SIDE EFFECT I H,ie 1 S is impolie l S is unappreciaive EXPLICIT [ 1 INITIAL ~ GOAL S will no aend T Declining MITIGATES S is polie S is accounable /\ S feels graiude for he offer Thanking S feels regre Apologizing Figure 4: A represenaion for Declining, Thanking, and Apologising no shown in his figure) Finally, here are miigaes finks beween nodes when wo effecs are conradicory A reasonable view of LeerGen's approach is ha here is a acr/p associaed wih each speech ac ha capures he causal chain of effecs ha poenlally follow from i While hese effecs could presumably be deermined by reasoning from firs principles, hese scrips can be viewed as sandard mehogs of achieving communicaive goals, and hey are essenially equivalen o he communicaive sraegies proposed by ohers (Marco, 1997) 42 Deermining Appropriae Speech Acs LeerGen's algorihm for producing a response involves 5 seps: 1 Mech he user's goal o one of he nodes (effecs) in he graph 2 From he maching effec, raverse graph finks "downward ~ oward he speech ac, checking he condiions on each link 3 For every pah ha reaches an ac by saisfying all condiions along he pah, add he ac o he new message by insaniaing he ac's ex emplae 4 Deec undesirable side effecs of each added speech ac by raversing all links back "up- ward ~ as far as possible 5 If an effec is indexed by a miigaes link, follow he link o he miigaing effec in he oher chain Coninue wih seps 2 and 3 As an example, consider he user's communicsive goal o make he hearer believe ha he speaker will no aend Leergen raverses he graph downwards o locae he speech ac Declining Afer deermining his speech ac, LeerGen hen raverses he graph upward, moving hrough is effecs, verifying ha none of hem conflic wih known speaker goals In his case, one of he effecs of Declining conflics wih he speaker's goal ha he hearer be lieves he speaker is polie A his poin, LeerGen generaes a new goal o miigae ha effec, and recursively uses is algorihm o locae speech acs o achieve ha goal Wih he failed goal of being perceived as polie, LeerGen's downward raversal locaes Thanking and Apologising as appropriae speech acs o miigae ha failure 43 An Alernaive To Planning This approach can be viewed as a form of reacive planning LeerGen can be viewed as having a simple goal (communicae a paricular belief o he hearer), forming a plan (finding a se of speech acs ha communicae his belief), analyzing he effecs of he plan (looking for user goals ha are violaed by hese speech acs), and opporunisically pursuing new goals (o miigae hese violaions) LeerGen differs significanly from mos oher effors in planning speech acs These effors explicily represen speech acs and heir effecs as plan operaors and aemp o synhesize sequences of operaors Unforunaely, as ohers have poined ou (Cohen and Levesque, 1980; 1990), plan operaors are no a good represenaion when acs have long chains of effecs Tha's because each chain ha resuis from a given ac mus be conflaed o a fia lis of effecs, or each effec mus be re-envisioned as an ac, wih one operaor for each effec and appropriae precondiions so he operaors can form he appropriae chain LeerGen's approach is mos similar o he alernaive o planning for speech-modeling proposed by Cohen and Levesque (1980, 1990) Their approach uses a se of inference rules and ac ype definiions and is explicily designed o capure sequences of his ype, cl c2 ci A(d) ---> El ---> E2 ---,---> Ei 1024
where A(d) is an ac ha communicaes proposiional conen d (definiional conen for some ac ype), which induces effec E1 under condiions cl, which induces effec E2 under condiions c2, and so on This rule formalism is direcly mappab]e o he condiionalised causal relaions used in our social perlocuions model, wih wo excepions One is ha we capure he rules wih an annoaed graph srucure ha makes he conneciviy among rules explici (scrips) The oher provide a specialized graph-raversal algorihm ha akes advanage of key properies of he graph, which allows us o subsiue e~cien graph raversal for generallsed planning 5 Implemenaion The curren implemenaion conains a very deailed model of speech ac effecs, conaining over 400 effecs and consrains I is able o generae a dozen differen ypes of messages, including iniiaing or erminaing a friendship, applying or resigning from a job, congraulaing or consoling someone, acceping or declining an inviaion, encouraging or discouraging someone from doing an ac, hanking someone, and apologizing o someone Each of hese differen message ypes includes an organizaional emplae ha places generaed acs in an appropriae order for he ask An imporan par of LeerGen is is ineracion wih he user Given a seleced message ype, LeerGen suggess a leas hree speech acs for he user o choose from For example, he hanking message ype (ie, make hem believe you feel graiude) can be insaniaed crediing (disribuing credi), offering (o repay), as well as an over expression of graiude (ie, hanking) For each ac chosen by he user, he sysem queries he user for he background informaion needed o insaniae an appropriae ex emplae 6 Limiaions and Fuure Work The model currenly has hree major limiaions Firs, i does no cover all aspecs of social ineracions For example, i does no have condiions or effecs involving he relaive saus of he speaker and hearer, or specialized roles hey migh play (eg, judge, employer, and so on) Second, he condiions on exacly when effecs occur need o be elabo- raed significanly Finally, here are socially-relaed speech acs we have no ye represened (eg, expressing sadness, joy, and so on) The primary implemenaion limiaion involves he background informaion required o deermine wheher various condiions hold Currenly, he implemenaion does no query he user for all he background informaion i could ake advanage of The reason is ha oo many queries makes he program loses is appeal as a work-saving device A relaed limiaion is ha is model of he speaker's goals is saic, raher han dynamic (eg, he speaker is always assumed o have a goal of being polie) We are addressing boh of hese problems by exploring echniques for forming a deailed user profile and applying across a large se of generaed leers The oher imporan limiaion is ha is organizaional and ex emplaes are no paricularly flexible (eg, hey demand a specific speech ac order and hey realize each speech ac as a single senence) One way o address his problem is o ake he se of speech acs ha LeerGen wans o generae as a goal and o plan exacly how hey will be realized (Hovy, 1993; Moore and Paris, 1994; Hobbs, 1982) One ineresing area for fuure exploraion is he problem of applying he model o leer undersanding as well as generaion This problem is poenially difllcul, as here are a variey of social reasons why a paricular speech ac migh have appeared For example, he hanking ac migh have been included in he example of Figure 1 in order o lessen he social deb he invies owes o he invier, or o avoid insuling he invier hrough curness, or o make he invies feel ha he is a polie person, or simply ou of habi 7 Conclusions This paper has presened a compuaional model of he sodal perlocuionary effecs of speech acs Our model exends previous formal modela of speech acs o ake ino accoun effecs involving emoions, impressions, and he inerpersonal relaionship beween he speaker and he hearer In doing so, we have inegraed earlier resuls from naural language generaion on speech acs, from communicaion sudies on communicaion sraegies, and from social psychology on how ineracions affec personaliy rais We have used his model o consruc a prooype program ha generaes leers ha mee social goals This ask is a key aspec of any generalpurpose, inelligen, personal assisan ha is in- 1025
volved in mediaing inerpersonal ineracion References [i] J F Allen, L K Schuber, O Ferguson, P Heeman, C H Hwang, T Kao, M Li gh, N G Marin, B W Miller, M Poesio, and D R Traum 1994 The TRAINS projec: A case sudl/ in building a eon~ersaiona! planning agen Technical Repor 532, Compuer Science Deparmen, Universiy of Rocheser, Rocheser, N'Y [2] D Appel 1985 Planning Engl~h senences New York: Cambridge Universiy Press [3] JL Ausin 1975 Ho~o o do faings ~oifh ~oorda Cambridge, MA:Harvard Universiy Press [4] D Barnlund 1989 Communicae sl/les of Japanese and Ame~canJ: Images and Realiies Belmon, CA: Wadsworh Publishing [5] B Bruce 1975 Belief spsema and language undersanding Technical Repor 2973 Cambridge, MA: Bol, Beranek, and Newman [6] D Chin 1989 KNOME: Modeling Wha The User Knows In UC In, A Kobsa and W Wah]ser (eds): User Modeling in Dialog glisens Berlin, Heidelberg: Spring-Verlag, pages 74-107 [7] P R Cohen and H J Levesque 1990 Raiohal ineracion as a bash for communicaion In Cohen, Morgan, and Pollack 1990 [8] P R Cohen and C R Perrsul 1979 Elemens of a plan-based heory of speech acs Cogniive Science 3 177-212 [13] [14] [15] [16] [17] [18] [19] [20] [211 M Ringie Hil Isdale, NJ: Lawrence Erlbaum Auocises E H Hovy 1988 (~'enerafing nau~ language under pragmaic eonarain~ Hlllsdale, NJ: Lawrence Erlbaum Associaes E H Hovy 1993 Auomaed discourse generaion using discourse srucure rela ions Ar~/~ j~cia! Inelligence 63 341-385 D Kiesler 1983 The 1982 inerpersonal cirde: A axonomy for complemenariy in human ransacions Psyclwlogica! Renew (90) 3 185-214 3 D Moore and C L Paris 1994 p]ann;ng ex for advisory dialogues: Capuring inenional and rheorical informaion Compu6- iona~ I, inguiaics (19) 4 651-694 A Orony, G Clore, and A Collins 1988 The cogniive s~rucuro of emoion~ New York: Cambridge Universiy Press D Pauler, 1999 A Compuer Mode! of Sraegic AspecfJ of Ineepersona! Communicaion, Forhcoming Phd Disseraion C R Perraul and 2 Allen 1980 A planbased analysis of indirec speech acs American Joul-aa! of Compuaional Linguiaics (6) 3-4 167-182 J Searle 1069 SpeecK ac~ New York: Cambridge Universiy Press A Wiersbicka 1987 Engliah speech ac ~erbs Sydney: Academic Press [9] P R Cohen, J Morgan, and M E Pollack (eds) 1990 Inenions in communicaion Cambridge, MA: MIT Press [10] J A Daly and 3 M Wiemann (eds) 1994 Sraegic inerpersonal communicaion Hillsdale, NJ: Lawrence Erlbaum Associaes [11] H P Grice 1975 Logic and conversaion In Syno, z and semanics III: Speech ac~, (eds) P Cole and 3 L Morgan New York: Academic Press [12] J R Hobbs 1982 Towards an undersanding of coherence in discourse In S ~raegies for na~uwi language processing, (eds) W Lehner and 1026