Causal Relation Extraction

Tí ol: Causal Rela ion Ex ac ion Volum: I Alumne: Edua do Blanco Villa Di ec o /Ponen : Dan Moldo an / Nu ia Cas ell A iño Depa amen : LSI Da a: 7 no emb e 2007 DADES DEL PROJECTE Tí ol del P ojec e: Causal Rela ion Ex ac ion Nom de l'es udian : Edua do Blanco Villa Ti ulació: Enginye ia en In o mà ica C èdi s: 37.5 Di ec o /Ponen : Dan Moldo an / Nu ia Cas ell A iño Depa amen : LSI MEMBRES DEL TRIBUNAL (nom i signa u a) P esiden : Vocal: Sec e a i: QUALIFICACIÓ Quali icació numè ica: Quali icació desc ip i a: Da a: Causal Rela ion Ex ac ion Edua do Blanco No embe 7, 2007 1 2 P e acio El p esen e documen o desc ibe el abajo hecho du an e la ealizaci´on del P oyec o Final de Ca e a (PFC), ´ul imo equisi o pa a ob ene los ´ı ulos de Enginye ia en In o m`a ica yM`as e Eu opeu en Llengua ge i Pa la po la Uni e si a Poli `ecnica de Ca alunya (UPC). El abajo ha sido ealizado en el Human Language Technology Resea ch Ins i u e (HLTRI), pe enecien e a The Uni e si y o Texas a Dallas (UTD). El di ec o es Dan Moldo an, Co-Di ec o del HLTRI. Nu ia Cas ell es la p o eso a de la UPC que me puso en con ac o con Dan Moldo an y sigui´o mis pasos en el HLTRI. La es ancia en el HLTRI se ex endi´o desde el 1 de eb e o al 31 de agos o de 2007, comp endiendo un o al de sie e meses. El PFC puede se cali icado como un abajo de in es igaci´on. Se han es udiado las elaciones sem´an icas de causalidad y se p opone un m´e odo pa a su de ecci´on y ex acci´on. No se p esen a un sis ema come cial o un p o o ipo lis o pa a su u ilizaci´on di ec a. Las a eas desa olladas han sido las ´ıpicas de es e ipo de abajos: lec u a de li e a u a elacionada con el campo y an´alisis del es ado del a e (2 meses), de inici´on de los obje i os y expe imen os a ealiza (1 mes), ealizaci´on de los expe imen os y ex acci´on de esul ados (3 meses), y po ´ul imo, esc i u a de la memo ia explicando el abajo ealizado (1 mes). 3 P e aci Aques documen desc iu el eball e du an la eali zaci´o del P ojec e Final de Ca e a (PFC), ´ul im equisi pe ob eni els ´ı ols d’Enginye ia en In o m`a ica i M`as e Eu opeu en Llengua ge i Pa la pe la Uni e si a Poli `ecnica de Ca alunya (UPC). La eina ha sigu eali zada en el Human Language Technology Resea ch Ins i u e (HLTRI), pe anyen a The Uni e si y o Texas a Dallas (UTD). El di ec o ´es Dan Moldo an, Co-Di ec o del HLTRI. Nu ia Cas ell ´es la p o esso a de la UPC que em a posa en con ac e amb Dan Moldo an i qui ha segui les me es passes en el HLTRI. L’es ada al HLTRI a comen¸ca l’1 de eb e de 2007 i a acaba el 31 d’agos del ma eix any. El PFC es po conside a un eball d’in es igaci´o. S’han es udia les elacions sem`an iques de causali a y es p oposa un m`e ode pe la se a de ecci´o i ex acci´o. No es p esen a un sis ema come cial ni un p o o ipus p epa a pe se u ili za di ec amen . Les asques desen olupades s´on les ´ıpiques d’un eball d’aques es ca ac e ´ıs iques: lec u a de la li e a u a elacionada amb el camp y an`alisi de l’es a de l’a (2 mesos), de ini i´o dels objec ius i expe imen s a eali za (1 mes), eali zaci´o dels ex- pe imen s i ex acci´o de esul a s (3 mesos), i pe ´ul im, esc ip u a de la mem`o ia explican la eina e a (1 mes). CONTENTS 4 Con en s 1 In oduc ion 6 2 Seman ic Rela ions 8 2.1 Examples o se s o seman ic ela ions . . . . . . . . . . . . . . . . 8 2.2 The exp ession o seman ic ela ions . . . . . . . . . . . . . . . . . 9 2.3 The di icul y o deciding he seman ic ela ion . . . . . . . . . . . 11 3 Causa ions 13 3.1 InPhilosophy ............................. 13 3.2 InPsychology............................. 14 3.3 In Theo e ical Linguis ics . . . . . . . . . . . . . . . . . . . . . . 15 3.3.1 In Typological Linguis ics . . . . . . . . . . . . . . . . . . 17 3.3.2 In he Cogni i e App oach . . . . . . . . . . . . . . . . . . 17 3.3.3 Wi hin Discou se Theo y . . . . . . . . . . . . . . . . . . . 18 3.4 Causa ion and o he seman ic ela ions . . . . . . . . . . . . . . . 18 3.5 Encoding o Causa ion . . . . . . . . . . . . . . . . . . . . . . . . 19 4 P e ious wo k 21 4.1 On Seman ic ela ions . . . . . . . . . . . . . . . . . . . . . . . . 21 4.2 Focused on causal ela ions . . . . . . . . . . . . . . . . . . . . . . 22 5 The me hod 25 5.1 Syn ac ic pa e ns ha encode causa ion . . . . . . . . . . . . . . 25 5.2 Pa e n ma ching . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 5.3 Machine Lea ning app oach . . . . . . . . . . . . . . . . . . . . . 30 5.3.1 Da a p epa a ion . . . . . . . . . . . . . . . . . . . . . . . 31 5.3.2 Fea u e selec ion . . . . . . . . . . . . . . . . . . . . . . . 31 5.3.3 Fea u e ex ac ion . . . . . . . . . . . . . . . . . . . . . . 35 5.3.4 Machine Lea ning algo i hm . . . . . . . . . . . . . . . . . 37 5.3.5 E alua ion........................... 39 5.3.6 E o analysis . . . . . . . . . . . . . . . . . . . . . . . . . 40 6 Limi a ions, Scope and Ex ensions 42 7 Conclusions and u he wo k 44 Re e ences 46 A Lis o POS ags 50 B Decision ees lea ned 51 B.1 Decision T ee no. 1: . . . . . . . . . . . . . . . . . . . . . . . . . . 51 B.2 Decision T ee no. 2: . . . . . . . . . . . . . . . . . . . . . . . . . . 51 CONTENTS 5 B.3 Decision T ee no. 3: . . . . . . . . . . . . . . . . . . . . . . . . . . 52 B.4 Decision T ee no. 4: . . . . . . . . . . . . . . . . . . . . . . . . . . 53 B.5 Decision T ee no. 5: . . . . . . . . . . . . . . . . . . . . . . . . . . 53 B.6 Decision T ee no. 6: . . . . . . . . . . . . . . . . . . . . . . . . . . 54 B.7 Decision T ee no. 7: . . . . . . . . . . . . . . . . . . . . . . . . . . 54 B.8 Decision T ee no. 8: . . . . . . . . . . . . . . . . . . . . . . . . . . 55 B.9 Decision T ee no. 9: . . . . . . . . . . . . . . . . . . . . . . . . . . 59 B.10 Decision T ee no. 10: . . . . . . . . . . . . . . . . . . . . . . . . . 59 2 SEMANTIC RELATIONS 12 o he o m [Noun Noun], 0.68 when classi ying complex nominals o he o m [Adjec i e Noun], and 0.65 when classi ying geni i es. [Chang and Choi, 2006] epo an ag eemen o 0.73 when classi ying manually a se o examples be ween only wo classes: encoding causa ion and no encoding causa ion. The low coe icien s gi e an idea o he di icul y o he ask: e en humans anno a o s o en disag ee when classi ying manually. 3 CAUSATIONS 13 3 Causa ions B oadly speaking, causa ion can be de ined as a ela ion be ween wo e en s: cause and e ec . Cause is he p oduce o he e ec , and e ec he esul o he cause. This de ini ion is ob iously ambiguous, and i doesn’ d aw a line be ween wha is a causa ion and wha is no . Causal ela ions ha e been s udied in se e al ields. Di e en heo ies o cau- sa ion coexis in Philosophy, Psychology and Linguis ics. 3.1 In Philosophy In his p esen a ion we ollow he ideas o [Whi e, 1990], which o e s summa y ske ches o he mos impo an poin s o each heo y. In Regula i y Theo ies [Hume, 1739], causa ion is seen as a cons an conjunc- ion be ween e en s, associa ed wi h p io i y in ime and con igui y in ime and, whe e ele an , space. The conjunc ion be ween e en s is hough o be impe - ec , inde e mina e, asymme ic and spu ious. The e o e, causa ion is de ined as a p obable conjunc ion o wo e en s and h ea ed using P obabili y Theo ies. The basic o m o he heo y says ha one e en causes ano he i i is ollowed by i and is such ha e en s o he i s kind a e egula ly ollowed by e en s o he second kind. Fo o he s, causa ion is a condi ion o he occu ence o an e en [Mill, 1843], [Sosa, 1975], [Mackie, 1980] [Suppes, 1970]: •A Su icien Condi ion CAUSE(e en 1, e en 2), i he exis ence o e en 1implies he exis ence o e en 2. •A Necessa y and Su icien Condi ion CAUSE(e en 1, e en 2), i he exis ence o e en 1implies he exis ence o e en 2and he exis ence o e en 2implies he exis ence o e en 1. •Insu icien bu Necessa y pa o an Unnecessa y bu Su icien (INUS) Condi ion CAUSE(e en 1, e en 2) i e en 1doesn’ imply he exis ence o e en 2, bu is pa o a mo e complica ed condi ion ha imply e en 2. A good example o unde s and INUS condi ions can be ound in [Hi ch- cock, 2007]: “. . . an INUS condi ion o some e ec is an insu icien bu non- edundan pa o an unnecessa y bu su icien condi ion. Suppose, o example, ha a li ma ch causes a o es i e. The ligh ing o he ma ch, by i sel , is no su icien ; many ma ches a e li wi hou ensuing o es i es. The li ma ch is, howe e , a pa o some cons ella ion o condi ions ha 3 CAUSATIONS 14 a e join ly su icien o he i e. Mo eo e , gi en ha his se o condi- ions occu ed, a he han some o he se su icien o i e, he ligh ing o he ma ch was necessa y: i es do no occu in such ci cums ances when li ma ches a e no p esen .” The Coun e ac ual App oach [Lewis, 1973a,Lewis, 1973b] sees causa ion as wha would ha e happened i some hing we e he case han in ac is no he case. In o he wo ds, e en 1causes e en 2jus in case i is ue ha i e en 1had no occu ed, hen e en 2would no ha e occu ed. Se e al esea che s c i icize his app oach and p opose di e en solu ions o he p oblems o he heo y. Fo example, [Mu ali, 1997] p oposes a solu ion o he p eemp ion p oblem, he p oblem o agile causes and inde e minis ic causa ion. 3.2 In Psychology Psychology is conce ned wi h how people unde s and and pe cei e causa ion, as well as how hey make causal in e ences and a ibu ions. Wha a e he conse- quences o hinking ha wo e en s a e causally ela ed is also s udied in Psy- chology. Di e en hypo hesis ha e been made abou he so o hings people may belie e can be causes [Whi e, 1990]: •E en s o happenings. •S anding condi ions o s a es o a ai s, such as causal powe s o ma e ial pa icula s. •In e ac ions be ween occu ences and s able p ope ies o hings. •Condi ions, such as necessa y and su icien condi ions o INUS condi ions. Psychologis s ha e pe o med se e al expe imen s in o de o explain empi ically he pe cep ion o causa ions by humans and he implica ions o conside ing wo e en s casually ela ed. [Glymou , 2003] o e s a o malism based on Bayes ne s which allows o ein e p e expe imen s on human judgmen , o e s a p ecise in e - p e a ion o mechanisms and allows gene aliza ions o exis ing heo ies o causal lea ning. [Eagleman and Holcombe, 2002] p esen an expe imen ha ela es causali y and he pe cep ion o ime. They no e ha i someone holds ou his hand and snaps his inge s, he will no no ice any di e ence in he ime ha he in ends o mo e his inge s and he ime he hea s he snap. Howe e , humans a e able o pe cei e a ound 25 miliseconds di e ences in iming and he audi o y signals co espond- ing o he snap we e being p ocessed by he ne ous sys em o mo e han 100 miliseconds. 3 CAUSATIONS 15 In a e y simple expe imen , subjec s olun a ily p essed a key ha caused a one o ollow. When asked o judge he ime o he keyp ess and he ime o he one, subjec s pe cei ed he keyp ess o occu la e and he one ea lie han i hese e en s had occu ed alone. The au ho s conclude ha when wo e en s a e casually ela ed, he pe cei ed ime o hese wo e en s shi s owa ds each o he . 3.3 In Theo e ical Linguis ics Causa ion has been in es iga ed by linguis ics unde a a ie y o opics. In his sec ion we will ocus on he wo k done on Typological Linguis ics, he Cogni i e App oach and Discou se Theo y. [S e anowi sch, 2001] o e s h ee p inciples o causa ion cons ual. These a e concep ual s a egies used o cons ue causal links be ween e en s and pa ici- pan s; hey a emp o iden i y wha in he human language hough leads o he pe cep ion o causal ela ions: 1. The Tempo al Succession P inciple Gi en wo e en s, Aand B, a concep ualize Cpe cei es Aas he cause o Bi and only i (a) A( egula ly) p ecedes B, and (b) Aand Ba e empo ally con iguous, and/o (c) Aand Ba e spa ially con iguous. 2. The Coun e ac uali y P inciple Gi en wo e en s, Aand B, a concep ualize Cpe cei es Aas he cause o Bi and only i (a) Aoccu s and Boccu s, and (b) (Cbelie es ha ) Bwould no ha e occu ed i Ahad no occu ed. 3. The T ansmission o Ene gy Gi en wo e en s, Aand B, a concep ualize Cpe cei es Aas he cause o Bi and only i (a) Cknows ha Acan ansmi ene gy, and (b) Cbelie es ha A ansmi ed ene gy o some pa o he eal wo ld and he eby gene a ed B. Two e en s a e conside ed o be causally ela ed i one o mo e o he h ee p in- ciples apply. [S e anowi sch, 2001] also o e s a se o a ibu es o causa ions e en s. These a ibu es help speci ying a causa ion and classi y hem: 3 CAUSATIONS 16 •The cause is he p oduce o he cause. We can cha ac e ize i using he ollowing se o p ope ies: –Anima e ∗in en ional o non in en ional ∗ oli ional o non oli ional ∗ he deg ee o he cause ’s con ol ∗whe he o no he cause is in a posi ion o au ho i y o e he causee –Inanima e ∗objec o e en No e ha in en ionali y en ails oli ionali y and con ol. •The causee is he p oduce o he e ec . We can cha ac e ize i using he ollowing se o p ope ies: –Anima e ∗ he deg ee o con ol ha he causee has o e he esul . ∗whe he o no he causee esis he cause ; i he causee esis he e migh be a ious mo i a ions o doing so. –Inanima e ∗objec o e en •The causing e en is he e en pe o med by he cause . We can cha ac- e ize i using he ollowing p ope ies: –Di ec i in ol es con ac be ween he cause and causee –Indi ec i i doesn’ in ol e con ac . •The esul ing e en is he e en p e o med by he causee. We can dis in- guish be ween: –immedia e o nonimmedia e, depending on he iming o he causing and esul ing e en s. I hey occu simul aneously wi h o he esul ing e en occu s igh a e he causing e en , i is immedia e; i a delay exis s be ween bo h e en s i is nonimmedia e. –s a ic o dynamic, depending on he na u e o he e en . 3 CAUSATIONS 17 3.3.1 In Typological Linguis ics [Shiba ani, 1976] p oposed se e al seman ic pa ame e s o causa i e cons uc- ions, including: •Coe ci e o noncoe ci e, depending on he deg ee o o ce ha he cause imposes on he causee. The cause may o ce,pe suade o gen ly sugges he causee. •Pe missi e o nonpe missi e. I pe missi e, we can subclassi y hem in ou ypes: –The cause o bea s (o omi s) p e en ion (o in e en ion). –The cause ac i ely gi es pe mission o he causee o do some hing. –The cause a emp s bu ails o p e en some hing om happening. –The cause gi es up and does no in e ene wi h he caused e en . •Di ec o indi ec , depending on he p e ious planing o he cause abou b inging a change. I di ec , we can dis inguish be ween di ec i e and ma- nipula i e. I he causee unc ions as a oli ional en i y, i is di ec i e, i he oli ion o he causee is absen hen is manipula i e. Acco ding o he au ho , his seman ic classi ica ion has been p o en ele an o many languages, and hus seems o be uni e sally applicable. 3.3.2 In he Cogni i e App oach Wi hin he cogni i e app oach one o he mos impo an heo ies is Fo ce Dy- namics [Talmy, 2000]. Fo ce Dynamics is a seman ic ca ego y co e ing how en i ies in e ac wi h espec o o ce. I s a ed ou as a gene aliza ion o e he adi ional no ion o causa i e, analyzing causa ion in o ine p imi i es and b inging he no ions o le ing,hinde ing, and helping in o he discussion. The heo y de ines concep s such as exe ion o o ce, esis ance o such exe ion and he o e coming o such esis ance,blockage o a o ce and he emo al o such blockage, e c. The basic concep ual p imi i es a e [Talmy, 2000]: •Two en i ies exe a o ce each on he o he . One is o ego unded o singled ou o ocal a en ion ( he agonis ); he o he is conside ed o he e ec i has on he agonis ( he an agonis ). •an en i y is aken o exe a o ce by i ue o an in insic endency owa ds mani es ing i , owa ds mo ion (ac ion) o owa ds es (inac ion). •opposed o ces ha e di e en ela i e s eng h, he en i y ha is able o mani es i s endency a he expense o i s opponen is he s onge . 3 CAUSATIONS 18 •acco ding o hei ela i e s eng hs, he opposing o ces yield a esul an , assessed only o he agonis , as ac ion o inac ion. A simple example may help unde s anding he basis o he heo y. “The woman ell because he loo wan slippe y.” would be conside ed as a causa ion by mos o he people. In he example, he agonis is he woman and he an ag- onis he loo . Bo h en i ies exe a o ce ( he woman (suposedly) didn’ wan o all; he loo , by being slippe y, makes people all); he esul is he woman alling, so he an agonis wins and he esul is an ac ion. The comple e heo y is much deepe and is conside ed one o he majo con- ibu ions in he s udy o g amma and seman ics. 3.3.3 Wi hin Discou se Theo y Discou se Theo y ies o de ine he seman ic ela ions ha hold be ween di e en discou se uni s. Thus, casusa ion is h ea ed as ano he seman ic ela ion. Sec ion 2.1 desc ibes di e en se s o seman ic ela ions p oposed so a and sec ion 3.4 s a es he di e ences be ween a causa ion and i s closes seman ic ela ions. 3.4 Causa ion and o he seman ic ela ions As we ha e al eady discussed, i ’s no always easy o decide wha seman ic e- la ion holds be ween wo ex spans. The limi s be ween causa ions and o he seman ic ela ions a e explained in his sec ion [Talmy, 2000], [Lako , 1987]. The closes seman ic ela ions o causa ion a e condi ion,consequence, eason and in luence. condi ion is a causa ion whose cause is hypo he ical (e.g. “I he we e hand- some, he would be ma ied.”). consequence is a causa ion whose e ec is in- di ec o unin ended (e.g. “His esigna ion caused eg e among all classes.”); eason is a causa ion o decision, belie , eeling o ac ing (e.g. “I wen because I hough i would be in e es ing.”). In his wo k we conside all o hem as cau- sa ion. The dis inc ion be ween in luence and causa ion is a ma e o deg ee. The e is a con inuum be ween o ally independen e en s and o ally casually ela ed e en s. An in luence holds be ween e en 1and e en 2i e en 1a ec s he manne o in ensi y o e en 2, bu does no a ec he occu ence o e en 2. I i a ec s he occu ence, hen i is a causa ion. Fo example, “Ta ge ing skin cance ela i es imp o es sc eening.” encodes an in luence. pu pose can be in e p e ed as an in ended causa ion, i.e., e en 1is he pu pose o e en 2i e en 1implies e en 2and he cause is willing he e en 2 o exis . We 3 CAUSATIONS 19 don’ conside pu pose in his wo k. As we ha e al eady men ioned, seman ic ela ions may o e lap, i.e., be ween wo ex spans mo e han one seman ic ela ion may hold. A clea o e lap exis s be ween causa ion and empo al ela ions. By de ini ion, he cause should always occu be o e he e ec , so i e en 1causes e en 2, hen e en 1should occu be o e han e en 2. We should make clea he dis inc ion be ween wo e en s causally ela ed and wo e en s co ela ed. The age o a human being and he size o his ocabula y ipically a e co ela ed: he olde , he mo e ocabula y. Howe e , we all know ha ge ing olde does no cause o ha e a bigge ocabula y. 3.5 Encoding o Causa ion The basic a gumen s o a causa ion a e he cause and he e ec . F om he poin o iew o de ec ing causa ions, he ollowing dis inc ions may be use ul: •Ma ked o unma ked A causa ion is ma ked i he e is a speci ic linguis ic uni ha signal he ela ion. Fo example: –[ma ked]: I bough his book because I ead a good e iew. –[unma ked]: Be ca e ul. I ’s uns able. Unma ked causa ions a e mo e di icul o de ec and usually i is no easy o decide whe he hey encode a causa ion o no . •Ambigui y I he causa ion is ma ked, he ma k can signal always a causa- ion o i can signal some imes a causa ion. The o me kind is unambiguous, he la e ambiguous. Fo example: –[unambiguous]: He died because o he ex en o his inju ies. –[ambiguous signaling causa ion]: The s o m p oduced se e al small o nadoes and gus nadoes. –[ambiguous no signaling causa ion]: The ac o y p oduces wo kinds o cable. •Explici o implici A causa ion is explici i bo h a gumen s a e p esen . I any o i s a gumen s a e missing, hen i is a implici causa ion. Fo example: –[explici ]: She was h own ou o he swanky Ho el Excelsio a e she had un naked h ough i s ma ble halls. 3 CAUSATIONS 20 –[implici ]: John killed Bob. Al hough i migh no be ob ious, he second example encodes a cau- sa ion. The e ec a gumen , Bob’s dea h, is no explici . In his wo k, we ocus in ma ked and explici causa ions. 4 PREVIOUS WORK 21 4 P e ious wo k In his sec ion we p esen he wo k al eady done in Compu a ional Linguis ics on he de ec ion and ex ac ion o seman ic ela ions and pa icula ly on causa ions. 4.1 On Seman ic ela ions One o he i s p oposed sys ems o pe o m seman ic analysis was Tanka, [Ba ke and Szpakowicz, 1995]. The sys em pe o ms seman ic analysis in h ee le els: Clause-Le el Rela ionships Analysis (seman ic ela ionships be ween ac s, e en s o s a es ep esen ed syn ac ically by synde ically connec ed ini e clauses), Case Analysis (be ween he main e b and i s syn ac ic a gumen s) and Noun-Modi ie Rela ionships Analysis (be ween he head noun o a noun ph ase and i s modi ie s). The Clause-Le el Rela ionships hey conside ha e been de ailed in sec ion 2.1. The sys em is based on a lexicon manually buil and a se o ules ha allows o decide he igh ela ionships o a pa icula pai o ex spans. The au ho s e alua e he sys em wi h a legal ex ; he as majo i y o ela ionships sugges ed a e co ec . The e alua ion is pe o med wi h 100 sen ences andomly selec ed, 94 sugges ions a e co ec and 2 w ong; he sys em canno make a decision o 4 sen ences and he e o e he sugges ion consis s o a se o possible ela ionships. Seman ic ela ions in Noun Ph ases ha e been s udied in [Moldo an e al., 2004]. Using he se o 35 seman ic ela ions desc ibed in able 1 o sec ion 2, hey com- pa e he pe o mance o h ee lea ning models: Nai e Bayes, Decision T ees and a new model, Seman ic Sca e ing. The bes esul s a e ob ained wi h Seman ic Sca e ing, he F-measu e a ies om 0.33 o 0.75 depending on he syn ac ic pa e n. As ea u es, hey only used he seman ic classes o he head and modi ie noun. In [Gi ju e al., 2004] he au ho s s udy he seman ic ela ions in Nominalized Noun Ph ases. They use a se o lexical, syn ac ic and seman ic ea u es. The bes esul s a e ob ained wi h Suppo Vec o Machines, he F-measu e a ies om 0.61 o 0.71 depending on he syn ac ic pa e n. O he app oaches wo k wi h P obabili y Theo y and he lexico-syn ac ic pa - e ns ha exp ess causa ions. These app oaches almos don’ need supe ision, bu equi e a la ge co pus. In [Tu ney, 2006] he au ho s de ine he pe inence o a pa e n o a wo d pai . When classi ying noun-modi ie pai s, hey epo a F-measu e o 0.50. They only conside i e seman ic ela ions: causali y,pa icipan ,quali y,spa ial and empo ali y. Thei sys em is unsupe ised and he only equi emen s a e a la ge co pus (a ound 5 ×1010 wo ds) and a se o wo ds pai s wi h hei co esponding labels (a leas 600 pai s in hei expe imen s). [Pan el and Pennacchio i, 2006] also wo k wi h p obabili ies and p opose wo new measu es: pa e n eliabili y and ins ance eliabili y. The key idea is ha in 5 THE METHOD 28 Pa e n Ma ching Syn ac ic pa e ns ha encode causa ion Co pus (SemCo 2.1) POS agging and Chunke Shallow pa sed sen ences Ma ches No ma ches Figu e 2: Flowcha o he pa e n ma ching. and e bs he Wo dNe 2.1 sense numbe . Howe e , all he e b o ms ha e he POS ag VB, which migh no be enough i we wan o ell he di e ence be ween di e en e b enses. In o de o ge he igh POS ag o all he e b o ms, we use a POS agge [Schmid, 1994]. While clus e ing he di e en examples in he syn ac ic pa e n (1), we disco e ed ha he mos common ela o s exp essing causa ion a e as,a e ,because and since. The e o e, we decided o ocus only in he causa ions ha : •a e ins ances o pa e n (1), and •a e signaled by he ela o as, since, a e o because. We ound 1068 sen ences ha sa is y bo h condi ions. The dis ibu ion o he ela o s is shown in able 5. Table 5: Dis ibu ion o he ela o s. Rela o Ca dinali y because 381 as 330 a e 228 since 129 To al 1068 The ollowing is a eal example con ained in he co pus SemCo 2.1. The o ma is he ollowing: i s column POS ag2, second lemma, hi d Wo dNe 2.1 sense numbe and ou h wo d o m: 2A lis o ags and he pa s o speech co esponding o hem can be ound in appendix A 5 THE METHOD 29 POS ag lemma wnsn wo d o m NN o dina y 1 O dina y NNP pe son 1 Williams VB say 1 said PRP he -1 he , , -1 , RB oo 2 oo , , -1 , VB was -1 was VB subjec 1 subjec ed VB o -1 o JJ anonymous 1 anonymous NN call 1 calls RB soon 1 soon IN a e -1 a e PRP he -1 he VB schedule 1 scheduled DT he -1 he NN elec ion 1 elec ion . . -1 . A e he POS agging o a ach he igh POS ag o e bs and he chunking, he esul is he ollowing: POS ag lemma wnsn wo d o m chunke NN o dina y 1 O dina y B-NP NNP pe son 1 Williams I-NP VBD say 1 said B-VP PRP he -1 he B-NP , , -1 , O RB oo 2 oo B-ADVP , , -1 , O VBD be -1 was B-VP VBN subjec 1 subjec ed I-VP IN o -1 o B-PP JJ anonymous 1 anonymous B-NP NN call 1 calls I-NP RB soon 1 soon B-ADVP IN a e -1 a e B-SBAR PRP he -1 he B-NP VBN schedule 1 scheduled B-VP DT he -1 he B-NP NN elec ion 1 elec ion I-NP . . -1 . O No e ha he example shown ma ches he pa e n we conside : O dina y Williams said he, oo, [was subjec ed]V P o anonymous calls soon [a e ] el [he [scheduled]V PC he elec ion]C. Using he o mula (2), he a gumen o de is i s e ec and hen cause, a gumen sO de =EC ∗middle =EC, so he e ec is encoded in V P and he cause in V PC. No e ha he ou ela o s chosen no always signal a causa ion: 5 THE METHOD 30 •Rachel s ayed on a e he doc o had gone. •He has li ed in many houses since he mo ed o he S a es. •She was ying o con ain he bi e ness o he oice as she enuncia ed he wo ds oo dis inc ly. 5.3 Machine Lea ning app oach The es o he asks in o de o c ea e a model o he de ec ion o causa ions a e ep esen ed in igu e 3. Ma ches Manual Clasi ica ion (causa ion / no causa ion) + ea u e selec ion Fea u e ex ac ion ML algo i hm E alua ion T aining da a E alua ion da a Model Classi ied da a Pe o mance Figu e 3: Flowcha o he lea ning p ocess and e alua ion. So a we ha e a se o sen ences ha a e likely o exp ess causa ion, since hei syn ac ic s uc u e is likely o encode one. Ou goal is o classi y hem in sen ences exp essing causa ion and sen ences no exp essing causa ion. In o de o do so, we decided o use Machine Lea ning echniques. The p oblem i s easily in he pa adigm o classi ica ion. F om now on, we will call ins ance he ma ches ob ained a e he syn ac ic pa - e n ma ching. Thus, an ins ance should be classi ied as cause i i encodes a causa ion, o ¬cause, i i doesn’ . Le us de ine Xas he se o all he ins ances and he se C={cause, ¬cause}. The unc ion de ined in (3) associa es each ins ance wi h i s igh classi ica ion. Ob iously is unknown a p io i. :X−→ C(3) 5 THE METHOD 31 The challenge o he Machine Lea ning pa adigm is o lea n he unc ion based on some examples. In o de o do so, he lea ning algo i hm akes as inpu o each example a ea u e ec o and i s igh classi ica ion. The algo i hm is able o c ea e a model ha is an app oxima ion o he unc ion . Once he model is lea ned, we need o es ima e i s pe o mance. In o de o do so, a new se o examples is needed. The pe o mance is measu ed conside ing he di e ences be ween he classi ica ion acco ding o he model ob ained and he igh classi ica ion. Typically he pe o mance is no pe ec ; i a ies a lo depending on he ask. Fo mally, o each ins ance xi∈X, we de ine he ec o Fi= (xi1, xi2. . . xim), co esponding o he ea u es o he ins ance xi.xij deno es he ea u e j o he ins ance i. We deno e he se o examples used o aining T and he se used o es ing T e.T and T e a e o he o m {F1c1, F2c2. . . Fmcm};ci∈C. Since we ha e a ini e numbe o ins ances, we need o di ide hem in he T and T e se s a he e y beginning. 5.3.1 Da a p epa a ion The da a p epa a ion basically consis ed on classi ying each ins ance as causa ion o no causa ion. Ou o he 1068 ins ances, 517 we e classi ied as cause and 551 as ¬cause3. This ask is e y slow and comple ely manual. A manual anno a o has o exam- ine each ins ance and decide i i encodes a causa ion o no . Fu he mo e, he ask is no as easy as i may seem. Some imes i is no clea whe he a sen ence encodes a causa ion o no ; ambigui y occu s mo e o en han expec ed. Conside he example “A e ou yea s o ha d wo k, she g adua ed”. One could a gue ha he cause o he g adua ion is he ac ha she wo ked ha d o ou yea s. Howe e , aking in o accoun only he in o ma ion con ained in he sen- ence we canno conclude ha a causa ion holds. 5.3.2 Fea u e selec ion A he same ime han he manual classi ica ion, a ea u e selec ion was pe o med. I is no ob ious wha kind o ea u es a e he good ones o ou p oblem. Typical ea u es o Machine Lea ning app oaches in he Na u al Language P ocessing ield a e a mix u e o lexical and syn ac ic; la ely some esea che s ha e achie ed g ea esul s conside ing seman ic ea u es oo. Following his app oach, we look o clues in he h ee le el o analysis ha de e mine good ea u es. 3This means he baseline o he classi ica ion ask is 0.516, since 51.6% o he ins ances belong o ¬cause 5 THE METHOD 32 I may be a gued ha he e o needed o come up wi h a good se o ea u es is oo much. I is de ini ely a lo o wo k o do, specially conside ing ha i is manual. Howe e , he ca e ul examina ion o he examples allows us o eally unde s and he na u e o causa ions. The se o ea u es ha we e de ec ed as po en ially good ea u es, hei a ionale, alues, de ec ion and examples a e de ailed below: 1. Rela o •Ra ionale: A ela o can signal always, ne e o some imes a causa- ion. Fu he mo e, di e en combina ion o ela o s and o he ea u es can signal causa ion always, ne e o some imes. •Values: because, as, a e , since. •De ec ion: a he same ime han he pa e n ma ching •Examples: –Rela o s unambiguously causal: because ∗[cause]Pa - ime a me s gene ally mus pay highe p ices o supplies han ull- ime a me s because hey buy in smalle quan i ies. ∗[cause]Leade ship is lacking in ou socie y because i has no legi ima e place o de elop. –Rela o s ambiguous: a e ,since,as ∗[¬cause]F ank and he had me abou wo yea s a e she had a i ed. ∗[cause]Ma y s ood o se e al momen s wi h his mou h hang- ing open oolishly a e i had happened. ∗[¬cause]They had been a lessons in he schoolhouse since hey e u ned om Ha pe s Fe y. ∗[cause]The child en a e sa is ied wi h smalle amoun s o ood since all o i is high in quali y. 2. Rela o Modi ie s •Ra ionale: causa ions can ha dly be signaled by a ela o modi ied by an ad e b o p eposi ion. •Values: POS ags. •De ec ion: a he same ime han he pa e n ma ching. •Examples: –ad e b + a e almos always signals a empo al ela ion, no a causa ion: 5 THE METHOD 33 ∗Tom B annon had caugh up wi h he ou i sho ly a e he Magui es joined i . ∗This was long a e Mo se had le he house. –as + p eposi ion can ha dly signal a causa ion: ∗. . . he el he was no ing i , as i i we e some hing he migh hink abou when . . . ∗Alex nodded o he maid as hough no hing unusual we e aking place and en e ed he doc o ’s oom. 3. Seman ic Class Cause Ve b •Ra ionale: only ce ain e bs can exp ess a cause. •Values: Wo dNe 2.1 seman ic class4. •De ec ion: a he same ime han he pa e n ma ching. •Examples: –I he ela o is a e and he cause e b seman ic class is be- -3 56, hen i is a empo al ela ion, no a causa ion: ∗We hea d him be o e he e e showed, and we hea d him yelling a e he was ou o sigh . 4. Cause Ve b is Po en ially Causal •Ra ionale: i a e b sense’s gloss con ains he wo ds cause o o change, o is subsumed by a e b sense ha con ains he wo ds cause o o change, hen is mo e likely o exp ess a cause 7. •Values: yes, no. •De ec ion: examine gloss cause e bs and i s subsume s •Examples: – ing- -1 is subsumed by sound- -2, which gloss is “cause o sound”. 5. Seman ic Class E ec Ve b •Ra ionale: only ce ain e bs can exp ess an e ec . •Values: Wo dNe 2.1 seman ic class. •De ec ion: a he same ime han he pa e n ma ching. 4By Seman ic Class we mean he mos common subsume o he e b in Wo dNe 2.1. Fo example, he seman ic class o umina e- -1 is ea - -2 5be- -3 should be eaded as he hi d meaning o he e b be 6The gloss o be- -3 is “occupy a ce ain posi ion o a ea; be somewhe e; ‘Whe e is my umb ella?’ ‘The oolshed is in he back’; ‘Wha is behind his beha io ?’ ” 7Acco ding wi h he de ini ion, we ound 7,370 ou o 24,890 e b senses po encially causal 5 THE METHOD 34 •Examples: – eac - -1’s seman ic class is ac - -1,allow- -10’s seman ic class is ac - -1 and ma ginalize- -1 ’s seman ic class is ac - -1. –I he ela o is a e and he e ec e b seman ic class is exp ess- -28, hen is no a causa ion: ∗“My name’s Gisele”, he blonde said a e she o de ed a Sco ch. 6. E ec Ve b is Po en ially Causal •Ra ionale: i a e b sense’s gloss con ains he wo ds cause o o change, o is subsumed by a e b sense ha con ains he wo ds cause o o change, hen is mo e likely o exp ess an e ec . •Values: yes, no. •De ec ion: examine gloss cause e bs and i s subsume s •Examples: –walk- -3 is subsumed by a el- -1, which gloss is “change loca- ion’. 7. Ve b Tense Cause and E ec Ve b •Ra ionale: depending on he ela o , some e b enses a e no likely o exp ess causa ion •Values: p esen , pas , u u e, pe ec i e, p og essi e, condi ional, obli- ga ion, possibili y, . . . •De ec ion: examine he POS ags o he cause and e ec e bs •Examples: –I he ela o is as o a e , he cause e b is no a copula (seman ic class di e en han be- -1 9), and he cause e b ense is p esen , hen is no a causa ion: ∗[¬cause]Hen ie a was disco e ing in he p ocess o w i ing, as he bo n w i e does, no me ely . . . ∗[¬cause]To play he gui a as he aspi es will de ou his . . . ∗[¬cause]The Thaye Schools o e s a yea o pos g adua e s udy in somewha he same way, a e a boy wins a B.S. in enginee ing. ∗[cause]. . . you mus o gi e me as I am so o ge ul. 8The gloss o exp ess- -2 is “a icula e; ei he e bally o wi h a c y, shou , o noise; ‘She exp essed he ange ’; ‘He u e ed a cu se’ ” 9The gloss o be- -1 is “ha e he quali y o being; (copula, used wi h an adjec i e o a p edica e noun); ‘John is ich’; ‘This is no a good answe ’ ” 5 THE METHOD 35 –I he ela o is as and he e ec e b is condi ional, hen is no a causa ion: ∗She wouldn’ go o New Yo k as Maude sugges ed . . . –I he e ec e b is p og essi e, hen is no a causa ion: ∗The bu den o his sec e was p essing down on him, as i was on Lieu enan Becks om and his six enlis ed men. ∗. . . said Juani a, holding he ace e y s ill, ying o con- ain he bi e ness o he oice as enuncia ed he wo ds oo dis inc ly. –i he ela o is as and he e ec e b exp ess obliga ion, hen is a causa ion: ∗You mus do ha as I helped you. –i he e ec e b es passi e, hen i is mo e likely o exp ess a causa ion: ∗. . . and hen Richa d was shocked as, all a once, lames sho ou om he sha p ea u es o . . . 8. Cause and E ec Ve b Same Tense •Ra ionale: i bo h he cause and e ec e b a e pas simple, i can ha dly be a causa ion, mos o he ime i is jus a empo al ela ion. •Values: yes, no. •De ec ion: cause and e ec e b same ense •Examples: –... ailed o lou ish in New England as i did in o he pa s o he coun y. Fle ched nodded as he lis ened o he ins uc ions and said he would . . . 5.3.3 Fea u e ex ac ion The ea u e ex ac ion is done ex ac ing he in o ma ion de ailed below om he ela o , he wo ds su ounding i , and bo h VPs. No e ha wi h he ou ela o s chosen, VP exp ess allways he e ec and he VP con ained in C (VPC) he cause. The inal se o ea u es a e he ollowing: • ela o = {since, because, as, a e } • ela o Le Modi ica ion = {POS ags} • ela o Righ Modi ica ion = {POS ags} •Seman icClassVCause = {Wo dNe 2.1 sense numbe } 5 THE METHOD 36 • e bCauseIsPo en iallyCausal = {yes, no} •Seman icClassVE ec = {Wo dNe 2.1 sense numbe } • e bE ec IsPo en iallyCausal = {yes, no} •Fo bo h Ve b Ph ases, he ollowing ea u es a e ex ac ed: –P esen = {yes, no} –Pas = {yes, no} –Modal = {condi ional, obliga ion, possibili y, u u e, no} –Pe ec i e = {yes, no} –P og essi e = {yes, no} –Passi e = {yes, no} We exempli y he ea u e ex ac ion wi h he sen ence p esen ed in sec ion 5.2, O dina y Williams said he, oo, [was subjec ed]V P o anonymous calls soon [a e ] el [he [scheduled]V PC he elec ion]C: ela o a e cP esen no eP esen no ela o Le Modi ica ion RB cPas yes ePas yes ela o Righ Modi ica ion PRP cModal no eModal no seman icClassVCause 1 cPe ec i e no ePe ec i e no e bCauseIsPo en iallyCausal yes cP og essi e no eP og essi e no seman icClassVE ec 1 cPassi e no ePassi e yes e bE ec IsPo en iallyCausal no In he example, ‘cP esen ’ means “cause e b ense is p esen ”, ‘eP esen ’ means “e ec e b ense is p esen ”, and so on. The ex ac ion o he non ob ious ea u es is depic ed below: •We can easily ge he main e b o a VP: i always co esponds o he las elemen o he VP [Qui k e al., 1985]. Since we ha e al eady iden i ied he VP encoding cause and e ec du ing he pa e n ma ching, we can easily ge hei seman ic classes looking o he in o ma ion con ained in he co pus. •Once we know he seman ic class, we only need o examine he gloss o know i any o he e bs is po en ially causal. •The boolean ea u es ha iden i y he ense o he VP a e de ined as ollows: –P esen : yes i he POS ags VBP o VPZ a e p esen in he VP; no o he wise. –Pas : yes i he POS ag VBD is p esen in he VP; no o he wise. 5 THE METHOD 37 –Modal: ∗condi ional i he lemma would is p esen in he VP, ∗obliga ion i he lemmas mus ,ough o should a e p esen in he VP, ∗possibili y i he lemmas can,could,may o migh a e p esen in he VP, ∗ u u e i he lemmas will o shall a e p esen in he VP, ∗no o he wise. –Pe ec i e: yes i he cons uc ion ha e + VBN (lemma ha e ollowed by pa iciple) is p esen ; no o he wise. –P og essi e: yes i he POS ag VBG is p esen in he VP; no o he wise. –Passi e: yes i he cons uc ion be + VBN (lemma be ollowed by pa iciple) is p esen ; no o he wise. No e ha di e en combina ion o hese boolean ea u es can cap u e enses like ‘pas and passi e’, e.g. was subjec ed, o ‘obliga ion, p esen and pe ec i e’, e.g. should ha e gone. 5.3.4 Machine Lea ning algo i hm We ied di e en Machine Lea ning algo i hms using 10- old c oss- alida ion. The bes esul s ob ained a e shown in able 6. They we e ob ained using Bag- ging wi h C4.5 ees. Table 6: Resul s ob ained using Bagging wi h C4.5 ees du ing aining. Class P ecision Recall F-Measu e cause 0.969 0.839 0.899 ¬cause 0.865 0.975 0.917 The measu es o e alua ing he pe o mance o he Machine Lea ning al- go i hm a e he mos commonly used. P ecision measu es he po ion o he assigned ca ego ies ha we e co ec ; Recall he po ion o he co ec ca ego ies ha we e assigned; F-measu e is he weigh ed ha monic mean o P ecision and Recall. The h ee measu es ange om 0 o 1. Fo mally, we can de ine he h ee measu es as ollows: Co ec = X Co ec = Y Assigned = X a b Assigned = Y c d 7 CONCLUSIONS AND FURTHER WORK 44 7 Conclusions and u he wo k We ha e p oposed a sys em which yields a high P ecision and Recall o he de ec ion o causa ions encoded by one o he mos common syn ac ic pa e ns exp essing a causa ion: [VP el C], [ el C, VP]. The sys em is ela i ely simple and is able o de ec causa ions encoded in an open domain ex . So a esea ch in causal ela ion de ec ion has ocused in causa ion exp essed wi h noun ph ases, e.g. “The [inciden ]NP1p o oked [widesp ead p o es ]NP2’. We ocused on he de ec ion o causa ion be ween a e b and a subo dina e clause. [Ba ke and Szpakowicz, 1995] also wo k wi h clauses, bu hei app oach is e y di e en . The sys em p oposed uses shallow syn ac ic pa sing. S a e o he a chunke s yield pe o mance o 95 %, so we can assume we de ec mos o he pa e n in- s ances p esen in he inpu . Howe e , chunke s ake as inpu POS agged ex , and POS agge s also make mis akes. A key elemen o eally see he po en ial o he sys em would be o in eg a e i wi h o he sys ems ha ex ac seman ic ela ions. By doing so we could expe imen wi h in e ence ules ha combine causa ion and o he seman ic e- la ions. Fo example, i e en 1causes e en 2and e en 3is subsumed by e en 1, we can conclude han e en 3causes e en 2; i e en 1causes e en 2and e en 2 en ails e en 3, hen e en 1causes e en 3. Ano he possible in e ence ule would exp ess he ansi i y p ope y o causa ions. An example may cla i y he las wo examples: •Being i ed makes me sno e. We all know ha in o de o sno e you need o be asleep. causa ion(being i ed, sno e)∧en ail(sno ing, sleeping)⇒causa ion(being i ed, sleep) •He died o cance . Cance is caused by smoking. causa ion(cance , dea h)∧causa ion(smoking, cance )⇒causa ion(smoking, dea h) No e ha he in e ence ule allows us o ex ac mo e causal knowledge, bu he causa ions ex ac ed a e indi ec . The disco e ing o new ules and he ali- da ion o he new ules seem o be a g ea challenge. Ano he possible ex ension would be o deal wi h causal chains and in ica e causal ela ions. A causal chain can be de ined as a sequence o e en s ha lead up o some inal e ec . Examples (1) and (2) exempli y a causal chain and a in ica e causal ela ion espec i ely. 7 CONCLUSIONS AND FURTHER WORK 45 1. A wo ks become a only when hey anscend he simple ac s o hei exis ence, and hey can do ha only when hey blend wi h he sensibili y o he iewe . ( hey (a wo ks) blend wi h he sensibili y o he iewe )⇒( hey (a wo ks) anscend he simple ac s o hei exis ence)⇒(a wo ks become a )10. 2. I is lined p ima ily by indus ial de elopmen s and conc e e-block walls because he cons an a ic and emissions do no make i an a ac i e neighbo hood. ((cons an a ic and emissions)⇒(no an a ac i e neighbo hood)) ⇒ (lined up by indus ial de elopmen s and conc e e-block walls). 10‘x ⇒y’ should be eaded as “x causes y” REFERENCES 46 Re e ences [Bake e al., 1998] Bake , C. F., Fillmo e, C. J., and Lowe, J. B. (1998). The Be keley F ameNe p ojec . In P oceedings o he Thi y-Six h Annual Mee ing o he Associa ion o Compu a ional Linguis ics and Se en een h In e na ional Con e ence on Compu a ional Linguis ics, pages 86–90, San F ancisco, CA, USA. Mo gan Kau mann Publishe s. [Ba ke and Szpakowicz, 1995] Ba ke , K. and Szpakowicz, S. (1995). In e ac i e seman ic analysis o clause-le el ela ionships. In P oceedings o he Second Con- e ence o he Paci ic Associa ion o Compu a ional Linguis ics (PACLING– 95), pages 22–30, B isbane, Aus alia. [Blahe a and Cha niak, 2000] Blahe a, D. and Cha niak, E. (2000). Asigning unc ion ags o pa sed ex . In P oceedings o he 1s Annual Mee ing o he No h Ame ican Chap e o Associa ion o Compu a ional Linguis ics (NAACL), Sea le, WA, USA. [Chang and Choi, 2006] Chang, D.-S. and Choi, K.-S. (2006). Inc emen al cue ph ase lea ning and boo s apping me hod o causali y ex ac ion using cue ph ase and wo d pai p obabili ies. In o ma ion P ocessing and Managemen : an In e na ional Jou nal, 42(3):662–678. [Eagleman and Holcombe, 2002] Eagleman, D. and Holcombe, A. (2002). Causal- i y and he pe cep ion o ime. T ends in Cogni i e Sciences, 6(8):323–325. [Ga cia, 1997] Ga cia, D. (1997). Coa is, an nlp sys em o loca e exp essions o ac ions connec ed by causali y links. In EKAW ’97: P oceedings o he 10 h Eu- opean Wo kshop on Knowledge Acquisi ion, Modeling and Managemen , pages 347–352, London, UK. Sp inge -Ve lag. [Gildea and Ju a sky, 2002] Gildea, D. and Ju a sky, D. (2002). Au oma ic label- ing o seman ic oles. Compu a ional Linguis ics, 28(3):245–288. [Gi ju, 2003] Gi ju, R. (2003). Au oma ic de ec ion o causal ela ions o ques- ion answe ing. In P oceedings o he 41s ACL, Wo kshop on Mul ilingual Summa iza ion and Ques ion Answe ing. [Gi ju e al., 2003] Gi ju, R., Badulescu, A., and Moldo an, D. (2003). Lea n- ing seman ic cons ain s o he au oma ic disco e y o pa -whole ela ions. In NAACL ’03: P oceedings o he 2003 Con e ence o he No h Ame ican Chap e o he Associa ion o Compu a ional Linguis ics on Human Language Technology, pages 1–8, Mo is own, NJ, USA. [Gi ju e al., 2006] Gi ju, R., Badulescu, A., and Moldo an, D. (2006). Au oma ic disco e y o pa -whole ela ions. Compu a ional Linguis ics, 32(1):83–135. REFERENCES 47 [Gi ju e al., 2004] Gi ju, R., Giuglea, A.-M., Ol eanu, M., Fo u, O., Bolohan, O., and Moldo an, D. (2004). Suppo ec o machines applied o he classi i- ca ion o seman ic ela ions in nominalized noun ph ases. In P oceedings o he HLT-NAACL Wo kshop on Compu a ional Lexical Seman ics. [Gi ju and Moldo an, 2002] Gi ju, R. and Moldo an, D. (2002). Mining answe s o causa ion ques ions. In P oceedings o he Ame ican Associa ion o A i i- cial In elligence (AAAI) - Sp ing Symposium. [Glymou , 2003] Glymou , C. (2003). Lea ning, p edic ion and causal bayes ne s. T ends in Cogni i e Sciences, (7):43–48. [Hi chcock, 2007] Hi chcock, C. (Fall 2007). P obabilis ic causa ion. In Zal a, E. N., edi o , The S an o d Encyclopedia o Philosophy. The Me aphysics Re- sea ch Lab. Cen e o he S udy o Language and In o ma ion. S an o d Uni- e si y. [Hobbs, 2005] Hobbs, J. R. (2005). Towa d a use ul concep o causali y o lexical seman ics. Jou nal o Seman ics, 22(2):181–209. [Hume, 1739] Hume, D. (1978, i s published anonymously in 1739). A ea ise o human na u e. Ox o d Uni e si y P ess. [Joskowiscz e al., 1989] Joskowiscz, L., Ksiezyk, T., and G ishman, R. (1989). Deep domain models o discou se analysis. In The annual AI sys ems in go - e nmen con e ence, pages 195–200. [Kaplan and Be y-Rogghe, 1991] Kaplan, R. M. and Be y-Rogghe, G. (1991). Knowledge-based acquisi ion o causal ela ionships in ex . Knowl. Acquis., 3(3):317–337. [Khoo e al., 2000] Khoo, C. S. G., Chan, S., and Niu, Y. (2000). Ex ac ing causal knowledge om a medical da abase using g aphical pa e ns. In ACL ’00: P oceedings o he 38 h Annual Mee ing on Associa ion o Compu a ional Linguis ics, pages 336–343, Mo is own, NJ, USA. Associa ion o Compu a- ional Linguis ics. [Kingsbu y e al., 2002] Kingsbu y, P., Palme , M., and Ma cus, M. (2002). Adding seman ic anno a ion o he penn eebank. In P oceedings o he Human Language Technology Con e ence (HLT’02). [Lako , 1987] Lako , G. (1987). Women, Fi e, and Dange ous Things. Uni e si y o Chicago P ess. [Lewis, 1973a] Lewis, D. (1973a). Causa ion. Jou nal o Philosophy, 70:556–567. [Lewis, 1973b] Lewis, D. (1973b). Coun e ac uals. Ox o d: Blackwell. REFERENCES 48 [Mackie, 1980] Mackie, J. L. (1980). The cemen o he Uni e se. A S udy o Causa ion. Ox o d Uni e si y P ess. [Ma cu and Echihabi, 2002] Ma cu, D. and Echihabi, A. (2002). An unsupe ised app oach o ecognizing discou se ela ions. P oceedings o he 40 h Annual Mee ing o he Associa ion o Compu a ional Linguis ics (ACL’02), pages 368– 375. [Mill, 1843] Mill, J. (1967, o iginal wo k published 1843). A sys em o logic a i- ocina i e and induc i e. Longman. [Moldo an e al., 2004] Moldo an, D., Badulescu, A., Ta u, M., An ohe, D., and Gi ju, R. (2004). Models o he seman ic classi ica ion o noun ph ases. P o- ceedings o he Compu a ional Lexical Seman ics Wo kshop a HLT-NAACL. [Mu ali, 1997] Mu ali, R. (1997). A coun e ac ual analysis o causa ion. Mind, (106):263–277. [Pan el and Pennacchio i, 2006] Pan el, P. and Pennacchio i, M. (2006). Esp esso: le e aging gene ic pa e ns o au oma ically ha es ing seman ic e- la ions. In ACL ’06: P oceedings o he 21s In e na ional Con e ence on Com- pu a ional Linguis ics and he 44 h annual mee ing o he ACL, pages 113–120, Mo is own, NJ, USA. [Quinlan, 1993] Quinlan, J. (1993). C4.5: P og ams o Machine Lea ning. Mo - gan Kau mann. [Quinlan, 1996] Quinlan, J. (1996). Bagging, boos ing and c4.5. In AAAI/IAAI, Vol. 1, pages 725–730. [Qui k e al., 1985] Qui k, R., G eenbaum, S., Leech, G., and S a ick, J. (1985). A Comp ehensi e G amma o he English Language. Longman. [Schmid, 1994] Schmid, H. (1994). P obabilis ic pa -o -speech agging using deci- sion ees. In In e na ional Con e ence on New Me hods in Language P ocessing, Manches e , UK. [Shiba ani, 1976] Shiba ani, M. (1976). The G amma o Causa i e Cons uc- ions: A Conspec us. Academic P ess. [Sosa, 1975] Sosa, E. (1975). Causa ions and condi ionals. Ox o d Uni e si y P ess. [S e anowi sch, 2001] S e anowi sch, A. (2001). Cons uc ion Causa ion: A Couns uc ional G amma App oach o Analy ic Causa i es. PhD hesis, Rice Uni e si y. REFERENCES 49 [Suppes, 1970] Suppes, P. (1970). A p obabilis ic heo y o causali y. No h- Holland Publishing Co. [Talmy, 2000] Talmy, L. (2000). Towa d a Cogni i e Seman ics. MIT P ess. [Ta u, 2005] Ta u, M. (2005). Au oma ic disco e y o in en ions in ex and i s applica ion o ques ion answe ing. In P oceedings o he ACL S uden Resea ch Wo kshop, Ann A bon, MI, USA. [Tu ney, 2006] Tu ney, P. D. (2006). Exp essing implici seman ic ela ions wi h- ou supe ision. In ACL ’06: P oceedings o he 21s In e na ional Con e ence on Compu a ional Linguis ics and he 44 h annual mee ing o he ACL, pages 313–320, Mo is own, NJ, USA. [Whi e, 1990] Whi e, P. (1990). Ideas abou causa ion in philosophy and psychol- ogy. Psychological Bulle in, 108(1):3–18. [Wi en and F ank, 2005] Wi en, I. and F ank, E. (2005). Da a Mining: P ac i- cal machine lea ning ools and echniques. Mo gan Kau mann. A LIST OF POS TAGS 50 A Lis o POS ags Tag Pa o speech CC Coo dina ing conjunc ion CD Ca dinal numbe DT De e mine EX Exis en ial he e FW Fo eign wo d IN P eposi ion o subo dina ing conjunc ion JJ Adjec i e JJR Adjec i e, compa a i e JJS Adjec i e, supe la i e LS Lis i em ma ke MD Modal NN Noun, singula o mass NNS Noun, plu al NNP P ope noun, singula NNPS P ope noun, plu al PDT P ede e mine POS Possessi e ending PRP Pe sonal p onoun PRP$ Possessi e p onoun RB Ad e b RBR Ad e b, compa a i e RBS Ad e b, supe la i e RP Pa icle SYM Symbol TO o UH In e jec ion VB Ve b, base o m VBD Ve b, pas ense VBG Ve b, ge und o p esen pa iciple VBN Ve b, pas pa iciple VBP Ve b, non-3 d pe son singula p esen VBZ Ve b, 3 d pe son singula p esen WDT Wh-de e mine WP Wh-p onoun WP$ Possessi e wh-p onoun WRB Wh-ad e b B DECISION TREES LEARNED 51 B Decision ees lea ned B.1 Decision T ee no. 1: ela o = a e : n ela o = as: n ela o = since | ela o Le Modi ica ion = yJJ: y | ela o Le Modi ica ion = yIN: y | ela o Le Modi ica ion = yJJR: y | ela o Le Modi ica ion = yVBZ: y | ela o Le Modi ica ion = yDT: y | ela o Le Modi ica ion = yTO: y | ela o Le Modi ica ion = yVBG: y | ela o Le Modi ica ion = yRB: n | ela o Le Modi ica ion = yVBP: n | ela o Le Modi ica ion = yVBD: y | ela o Le Modi ica ion = yPRP: y | ela o Le Modi ica ion = yCD: y | ela o Le Modi ica ion = yVBN | | ePe ec i e = yes: n | | ePe ec i e = no | | | lexicalClue = yes: n | | | lexicalClue = no: y | ela o Le Modi ica ion = yCC: n | ela o Le Modi ica ion = yVB: y | ela o Le Modi ica ion = yNNP: y | ela o Le Modi ica ion = yNNS: n | ela o Le Modi ica ion = yNN | | ela o Righ Modi ica ion = yPP: n | | ela o Righ Modi ica ion = yJJ: y | | ela o Righ Modi ica ion = y*: n | | ela o Righ Modi ica ion = yIN: y | | ela o Righ Modi ica ion = yVBZ: n | | ela o Righ Modi ica ion = yDT | | | e bCausePo Causal = yes: y | | | e bCausePo Causal = no: n | | ela o Righ Modi ica ion = yTO: n | | ela o Righ Modi ica ion = yVBG: n | | ela o Righ Modi ica ion = yRB: y | | ela o Righ Modi ica ion = yWDT: n | | ela o Righ Modi ica ion = yVBP: n | | ela o Righ Modi ica ion = yVBD: n | | ela o Righ Modi ica ion = yPRP: n | | ela o Righ Modi ica ion = yVBN: n | | ela o Righ Modi ica ion = yCD: n | | ela o Righ Modi ica ion = yWRB: n | | ela o Righ Modi ica ion = yVB: n | | ela o Righ Modi ica ion = yNNP: n | | ela o Righ Modi ica ion = yNNS: n | | ela o Righ Modi ica ion = yEX: n | | ela o Righ Modi ica ion = yMD: n | | ela o Righ Modi ica ion = yPDT: n | | ela o Righ Modi ica ion = yNN: n | | ela o Righ Modi ica ion = yWP: n | | ela o Righ Modi ica ion = yPUNCT: n | ela o Le Modi ica ion = yPUNCT: y ela o = because: y Numbe o Lea es : 49 Size o he ee : 55 B.2 Decision T ee no. 2: ela o = a e : n B DECISION TREES LEARNED 52 ela o = as: n ela o = since | ePe ec i e = yes: n | ePe ec i e = no | | ela o Le Modi ica ion = yJJ: y | | ela o Le Modi ica ion = yIN: y | | ela o Le Modi ica ion = yJJR: y | | ela o Le Modi ica ion = yVBZ: y | | ela o Le Modi ica ion = yDT: y | | ela o Le Modi ica ion = yTO: y | | ela o Le Modi ica ion = yVBG: y | | ela o Le Modi ica ion = yRB: y | | ela o Le Modi ica ion = yVBP: n | | ela o Le Modi ica ion = yVBD: y | | ela o Le Modi ica ion = yPRP: y | | ela o Le Modi ica ion = yCD: y | | ela o Le Modi ica ion = yVBN: y | | ela o Le Modi ica ion = yCC: n | | ela o Le Modi ica ion = yVB: y | | ela o Le Modi ica ion = yNNP: y | | ela o Le Modi ica ion = yNNS: n | | ela o Le Modi ica ion = yNN | | | cPas = yes: n | | | cPas = no: y | | ela o Le Modi ica ion = yPUNCT: y ela o = because: y Numbe o Lea es : 24 Size o he ee : 28 B.3 Decision T ee no. 3: ela o = a e : n ela o = as: n ela o = since | ePe ec i e = yes: n | ePe ec i e = no | | ela o Le Modi ica ion = yJJ: y | | ela o Le Modi ica ion = yIN: n | | ela o Le Modi ica ion = yJJR: y | | ela o Le Modi ica ion = yVBZ: y | | ela o Le Modi ica ion = yDT: y | | ela o Le Modi ica ion = yTO: y | | ela o Le Modi ica ion = yVBG: y | | ela o Le Modi ica ion = yRB: n | | ela o Le Modi ica ion = yVBP: n | | ela o Le Modi ica ion = yVBD: y | | ela o Le Modi ica ion = yPRP: y | | ela o Le Modi ica ion = yCD: y | | ela o Le Modi ica ion = yVBN | | | ela o Righ Modi ica ion = yPP: y | | | ela o Righ Modi ica ion = yJJ: y | | | ela o Righ Modi ica ion = y*: y | | | ela o Righ Modi ica ion = yIN: y | | | ela o Righ Modi ica ion = yVBZ: y | | | ela o Righ Modi ica ion = yDT: y | | | ela o Righ Modi ica ion = yTO: y | | | ela o Righ Modi ica ion = yVBG: y | | | ela o Righ Modi ica ion = yRB: y | | | ela o Righ Modi ica ion = yWDT: y | | | ela o Righ Modi ica ion = yVBP: y | | | ela o Righ Modi ica ion = yVBD: y | | | ela o Righ Modi ica ion = yPRP: y | | | ela o Righ Modi ica ion = yVBN: y | | | ela o Righ Modi ica ion = yCD: y B DECISION TREES LEARNED 53 | | | ela o Righ Modi ica ion = yWRB: y | | | ela o Righ Modi ica ion = yVB: n | | | ela o Righ Modi ica ion = yNNP: y | | | ela o Righ Modi ica ion = yNNS: y | | | ela o Righ Modi ica ion = yEX: y | | | ela o Righ Modi ica ion = yMD: y | | | ela o Righ Modi ica ion = yPDT: y | | | ela o Righ Modi ica ion = yNN: y | | | ela o Righ Modi ica ion = yWP: y | | | ela o Righ Modi ica ion = yPUNCT: y | | ela o Le Modi ica ion = yCC: n | | ela o Le Modi ica ion = yVB: y | | ela o Le Modi ica ion = yNNP: y | | ela o Le Modi ica ion = yNNS: n | | ela o Le Modi ica ion = yNN | | | cPas = yes: n | | | cPas = no: y | | ela o Le Modi ica ion = yPUNCT: y ela o = because: y Numbe o Lea es : 48 Size o he ee : 53 B.4 Decision T ee no. 4: ela o = a e : n ela o = as: n ela o = since | ePe ec i e = yes: n | ePe ec i e = no | | ela o Le Modi ica ion = yJJ | | | lexicalClue = yes: n | | | lexicalClue = no: y | | ela o Le Modi ica ion = yIN: n | | ela o Le Modi ica ion = yJJR: y | | ela o Le Modi ica ion = yVBZ: y | | ela o Le Modi ica ion = yDT: y | | ela o Le Modi ica ion = yTO: y | | ela o Le Modi ica ion = yVBG: y | | ela o Le Modi ica ion = yRB | | | cPas = yes: n | | | cPas = no: y | | ela o Le Modi ica ion = yVBP: n | | ela o Le Modi ica ion = yVBD: y | | ela o Le Modi ica ion = yPRP: y | | ela o Le Modi ica ion = yCD: y | | ela o Le Modi ica ion = yVBN | | | lexicalClue = yes: n | | | lexicalClue = no: y | | ela o Le Modi ica ion = yCC: n | | ela o Le Modi ica ion = yVB: y | | ela o Le Modi ica ion = yNNP: y | | ela o Le Modi ica ion = yNNS: n | | ela o Le Modi ica ion = yNN | | | cPas = yes: n | | | cPas = no: y | | ela o Le Modi ica ion = yPUNCT: y ela o = because: y Numbe o Lea es : 27 Size o he ee : 34 B.5 Decision T ee no. 5: ela o = a e : n B DECISION TREES LEARNED 60 | | ela o Le Modi ica ion = yCC: n | | ela o Le Modi ica ion = yVB: y | | ela o Le Modi ica ion = yNNP: y | | ela o Le Modi ica ion = yNNS: n | | ela o Le Modi ica ion = yNN | | | cPas = yes | | | | ela o Righ Modi ica ion = yPP: n | | | | ela o Righ Modi ica ion = yJJ: n | | | | ela o Righ Modi ica ion = y*: n | | | | ela o Righ Modi ica ion = yIN: n | | | | ela o Righ Modi ica ion = yVBZ: n | | | | ela o Righ Modi ica ion = yDT: y | | | | ela o Righ Modi ica ion = yTO: n | | | | ela o Righ Modi ica ion = yVBG: n | | | | ela o Righ Modi ica ion = yRB: y | | | | ela o Righ Modi ica ion = yWDT: n | | | | ela o Righ Modi ica ion = yVBP: n | | | | ela o Righ Modi ica ion = yVBD: n | | | | ela o Righ Modi ica ion = yPRP: n | | | | ela o Righ Modi ica ion = yVBN: n | | | | ela o Righ Modi ica ion = yCD: n | | | | ela o Righ Modi ica ion = yWRB: n | | | | ela o Righ Modi ica ion = yVB: n | | | | ela o Righ Modi ica ion = yNNP: n | | | | ela o Righ Modi ica ion = yNNS: n | | | | ela o Righ Modi ica ion = yEX: n | | | | ela o Righ Modi ica ion = yMD: n | | | | ela o Righ Modi ica ion = yPDT: n | | | | ela o Righ Modi ica ion = yNN: n | | | | ela o Righ Modi ica ion = yWP: n | | | | ela o Righ Modi ica ion = yPUNCT: n | | | cPas = no: y | | ela o Le Modi ica ion = yPUNCT: y ela o = because: y Numbe o Lea es : 48 Size o he ee : 53