scieee Open visual document viewer

Extracting Lexical Units for Identifying Specialized Semantic Frames

Ana Ostroški Anić; Maja Lončar; Martina Pavić

Abstract

    The paper discusses the idea of using parallel corpora for the extraction of terminological units used to identify semantic frames of aviation, as the first step in developing a method for determining specialized semantic frames. The process of compiling the parallel corpus of texts taken from the Eur-Lex database using Sketch Engine tools is described first. The methodology of extracting term candidates and their manual verification is then presented, as well as the process of extracting and verifying relevant verbs. The results of term verification are discussed, with the focus on the role of verbs and verbal collocations in identifying relevant semantic frames and frame elements.

Full text

7 3Te minologija | 2019 | 26 doi.o g/10.35321/ e m26-04 Ex ac ing Lexical Uni s o Iden i ying Specialized Seman ic F ames1 AnA Os Oški Anić Ins i u e o C oa ian Language and Linguis ics MAjA LOnčA Ins i u e o C oa ian Language and Linguis ics MA inA PA ić Ins i u e o C oa ian Language and Linguis ics AbS Ac he pape discusses he idea o using pa allel co po a o he ex ac ion o e - minological uni s used o iden i y seman ic ames o a ia ion, as he i s s ep in de eloping a me hod o de e mining specialized seman ic ames. he p o- cess o compiling he pa allel co pus o ex s aken om he Eu -Lex da abase using Ske ch Engine ools is desc ibed i s . he me hodology o ex ac ing e m candida es and hei manual e i ica ion is hen p esen ed, as well as he p ocess o ex ac ing and e i ying ele an e bs. he esul s o e m e i i- ca ion a e discussed, wi h he ocus on he ole o e bs and e bal collo- ca ions in iden i ying ele an seman ic ames and ame elemen s. Keywo ds: a ia ion e minology, English-c oa ian co pus, seman ic ames, specialized knowledge, e minology. Ano AcIjA S aipsnyje s a s oma, kaip lygiag a ieji eks ynai gali bū i naudojami e minijos iene ams, aikomiems a iacijos seman iniams eimams nus a y i, a ink i. ai pi moji me odo specializuo iems seman iniams eimams nus a y i kū imo pako- pa. Pi miausia ap ašoma, kaip, emian is Eu -Lex duomenų baze i naudojan is Ske ch Engine p iemonėmis, engiamas lygiag a usis eks ų inkinys. Tada pa ei- kiami e minų a ian ų iš inkimo i jų pa ik inimo ankiniu būdu me odai bei 1 his pape is based on he alk “F om pa allel co po a o specialized seman ic ames”, gi en a he 3 d in e na ional con e ence on e minology Scien i ic, Adminis a i e and Educa ional Dimensions o Te min ol- ogy in Vilnius, 17–18 oc obe 2019. his wo k has been ully suppo ed by he c oa ian Science Foun- da ion unde he p ojec H ZZ-UIP-2017-05-7169. 7 4 Ana Os oški Anić Ex ac ing Lexical Uni s o Iden i ying Maja Lonča Specialized Seman ic F ames Ma ina Pa ić inkamų eiksmažodžių iš inkimo i ik inimo p ocesas. Ap a ian e minų pa ik inimo ezul a us ypa ingas dėmesys ski iamas eiksmažodžių i žodžių jungi- nių aidmeniui, nus a omi inkami seman iniai eimai i jų elemen ai. esminiai žodžiai: a iacijos e minija, anglų i k oa ų kalbų eks ynas, seman iniai eimai, specializuo os žinios, e minologija. In oDUc Ion he e ha e been nume ous applica ions o he heo y o F ame Seman- ics (Fillmo e 1976, 1982) in lexicog aphy and compu a ional linguis ics, mos known o which is ce ainly he lexical knowledge da abase F amene (Fillmo e e al. 2003; uppenho e e al. 2017). he wo k on mapping ela ions be ween syn ax and seman ics, which F amene essen ially is, has no only inspi ed simila p ojec s conce ning o he languages besides English, bu has also pa ed he way o esea ch on specialized knowledge and he o ganiza ion o specialized knowledge ca ego ies based on seman- ic ames. ega dless o whe he hey adhe e o F amene ’s me hodol- ogy closely o hey apply a modi ied app oach o iden i ying and delin- ea ing seman ic ames as ypes o ca ego ies, he p ojec s c ea ed a ound building esou ces designed o specialized knowledge communi ies need- ed o ake in o accoun he speci ics o e minology and he con ex o languages o speci ic pu poses as opposed o gene al language and i s use (Fabe e al. 2006; 2009; L’Homme 2012). he e has been no lexical esou ce de eloped o he c oa ian language ha is comple ely based on F amene ’s me hodology and seman ic in en- o y. ce ain seman ic ames ha e been de ined as pa o a la ge da a- base o concep ual me apho s in c oa ian (Despo e al. 2019). Howe e , a da abase o seman ic ames o a ia ion, which is being de eloped wi h- in he esea ch p ojec he Dynamici y o Specialized Knowledge Ca ego ies (DIKA, ihjj.h /dika), is he i s a emp a building such a specialized da abase o c oa ian. he p ojec deals wi h he desc ip ion o concep- ual and linguis ic le els o specialized knowledge ca ego ies wi hin he dynamic na u e o seman ic ames. he p ojec ’s key ou pu is a mul i- lingual e minological da abase in which a ia ion e minology is de ined in seman ic ames, including ame elemen s, concep ual ela ions, and igu a i e and ph aseological e minological uni s. A pa allel co pus o English and c oa ian ex s in he domain o ai a ic has been compiled o he pu poses o seman ic and syn ac ic analysis, and i will se e as 7 5Te minologija | 2019 | 26 he s a ing poin o e m ex ac ion and e m analysis. he subdomain o ai a ic has been chosen as he mos ep esen a i e aspec o he as and in e disciplina y ield o a ia ion. his pape discusses he idea o using pa allel co po a o he ex ac ion o e minological uni s e e ing o speci ic seman ic ames o a ia ion, as he i s s ep in de eloping a me hod o de e mining seman ic ames and hei concep ual ela ions in he ield o a ia ion. Al hough pa allel co po a ha e long been ecognized as in aluable sou ces o e m candida es o build- ing a ious specialized esou ces (Vin a 2000), hey ha e so a been li le used in applied esea ch e e ing o F ame Seman ics.2 he p ocess o compiling he pa allel co pus o ex s aken om he Eu -Lex da abase us- ing Ske ch Engine ools is desc ibed i s . he me hodology o ex ac ing e m candida es and hei manual e i ica ion is hen p esen ed, as well as he p ocess o ex ac ing and e i ying ele an e bs. he esul s o e m e i ica ion a e p esen ed and commen ed on, ollowed by he discussion o he po en ial use o e bs and e bal colloca ions in iden i ying ele an seman ic ames and ame elemen s in he ield o a ia ion.3 1. HEo E IcAL bAcKG oUnD In con empo a y e minology and LSP communi ies, specialized knowl- edge esou ces a e no longe buil a ound he desc ip ion o hie a chical ela ions only because i is conside ed hey do no e lec he dynamic na u e o he knowledge ca ego ies (Fabe e al. 2006; 2009; L’Homme, obichaud 2014). Mo eo e , he o mal dis inc ion be ween a dic iona y and a lexicon ha he whole adi ional e minog aphy is based on seems o ha e been blu ed wi h mo e and mo e specialized esou ces employing he me hods o compila ion and knowledge p esen a ion o he wise used in c ea ing gene al language esou ces (L’Homme e al. 2016; L’Homme 2018). Lexical and seman ic esou ces buil on he p emises o F ame Seman ics de ine knowledge in e ms o seman ic ames as a ype o dynamic ca ego ies o in ica e elemen s and mu ual ela ions. Each seman ic ame is de ined as a ype o e en o s a e wi h i s pa - icipan s, called ame elemen s (FEs). Seman ic ames a e in oked by 2 Pa allel co po a would ce ainly each hei ull po en ial in he c ea ion o mul ilingual lexical eposi o- ies based on seman ic ames (boas 2005). 3 We hank he anonymous e iewe s o hei in aluable commen s and sugges ions. 7 6 Ana Os oški Anić Ex ac ing Lexical Uni s o Iden i ying Maja Lonča Specialized Seman ic F ames Ma ina Pa ić lexical uni s, i.e. wo ds in e e yday language o e ms in specialized knowl- edge, which ins an ia e ame elemen s. he basic c i e ion o delimi ing one ame om ano he is ha all lexical uni s should “e oke he same ype o e en and sha e he same in en o y and con igu a ion o FEs” ( uppenho e e al. 2017: 384–385). his me hod o a i ing a he scope o seman ic ames and hei de ini ions could be, howe e , conside ed an a bi a y one because i does no equi e a p e ious “o e iew” o he a ailable lexical da a. Since many ames ollow a scena io-like composi- ion, hei desc ip ion usually s a s om selec ing ele an e bs ( up- penho e e al. 2010) o which app op ia e a gumen s a e added ollowing he anno a ion o examples ex ac ed om he co pus. Al hough lexical uni s and ame elemen s a e de ined acco ding o he analysed co pus da a, seman ic ames a e ne e heless mos ly in ui i ely iden i ied and named. In iden i ying seman ic ames o a ce ain specialized knowledge domain, i would be mo e sui able o i s delinea e he scope o knowledge being de ined, which in e minology wo k is done h ough he iden i ica- ion o key concep s and concep ual ela ions in he domain. e minological esea ch has closely kep ack o de elopmen s in lin- guis ics o e he pas wo decades, and has shi ed i s heo e ical ocus o analysing p ocesses and e en s as mo e dynamic specialized knowledge s uc u es. he e o e, i has become clea ha p edica i e e ms ( e bs, adjec i es, ad e bs and ce ain nouns) o en deno e key concep s o a domain e minology (L’Homme 1998; Pimen el 2012; L’Homme, o- bichaud 2014). Fo ins ance, e ms communica e, ly, ake o , and ans- po a e ypically p esen in he ield o a ia ion, as well as e ms o a numbe o p ocedu es ha we on ologically de ine as p ocesses o ac- i i ies. Al hough he no ion o e mhood is a he well unde s ood among e minologis s when i comes o de e mining he concep ual le el o e ms, he c i e ia o choosing which ypes o lexical uni s should be included as e ms in a e minological esou ce s ill di e depending on he ield analysed and he heo e ical amewo k applied. he c i e ia o de ining ypes o e bs ca ying specialized meaning p oposed by L’Homme (1998, 2015) a e mos e e ed o in e minological li e a u e, along wi h Lo en e’s c i e ia and he ypology o e bs in specialized discou se (2002, 2007). L’Homme dis inguishes among h ee ypes o e bs in specialized ex s: e bs speci ic o a ce ain ield (e.g. land, ake o ), e bs ha acqui e specialized meaning in he ield (e.g. ope a e), and e bs o gene al lan- 7 7Te minologija | 2019 | 26 guage.4 Ve bs ha belong o he i s and second ypes can be labelled as e ms, bu e bs o he second ype p esen a mo e in e es ing g oup o e minological analysis because he e a e cases when i is no clea whe h- e he meaning is di e en enough om he e b’s canonical o de aul gene al dic iona y meaning. he c i e ia o deciding whe he a e b akes on a new, specialized meaning in a gi en con ex should ye be mo e p ecisely de ined. he c i e ia o deciding whe he a e b can be classi ied as a e m depend, acco ding o L’Homme (1998, 2015), on i s a gumen s and i s mo pho- logical and seman ical ela ion o lexical uni s o he same meaning ha a e de ined as e ms. In sho , i a e b has a gumen s ha a e de ined as e ms, i is likely o be a e m i sel , as in he sen ence The pilo lies he helicop e . Also, i he noun akeo is de ined as a e m, hen he e b ake o should also be de ined as one. We ha e applied hese c i e ia in ou analysis o ex ac ed e bs, o which mo e is said in sec ion 4. 2. ME HoDS In o de o de elop a alid me hodology o iden i ying specialized se- man ic ames, se e al s eps had o be aken. Fi s , a pa allel English- c oa ian co pus was c ea ed. he second s age consis ed o e m ex ac ion and alida ion o e m candida es’ lis s. he hi d s ep included ex ac ing a lis o e bs and analysing hei colloca ions bea ing specialized meaning. 2.1. Co pus compila ion he co pus is compiled om he Di ec o y o legal ac s o he Eu o- pean Union, chap e T anspo policy, subchap e Ai anspo in English and c oa ian. ou o 220 documen s om he Ai anspo subchap e , 178 legal ac s a e aken ha ing bo h (English and c oa ian) language e sions. he ex s a e downloaded om he EU -Lex da abase,5 and en e ed in o he Ske ch Engine’s co pus compila ion module.6 he sim- ples way o c ea e a pa allel co pus in Ske ch Engine is o upload da a 4 L’Homme (2015) u he classi ies gene al language e bs in o suppo e bs (a e Le a 2002) and discu si e e bs ( ollowing Lo en e and he classi ica ion om 2002). See Pimen el (2017) o a ecen elabo a ion o hese wo app oaches o ypology o e bs in specialized ex s. 5 A ailable a : h ps://eu -lex.eu opa.eu/b owse/di ec o ies/legisla ion.h ml [accessed 05.09.2019]. 6 A ailable a : h ps://www.ske chengine.eu/use -guide/use -manual/co po a/se ing-up-pa allel-co po a/ [accessed 05.09.2019]. 7 8 Ana Os oški Anić Ex ac ing Lexical Uni s o Iden i ying Maja Lonča Specialized Seman ic F ames Ma ina Pa ić in a abula o ma such as a sp eadshee (Excel). Sp eadshee s mus con ain language names in he i s ow and hen aligned segmen s (wo d, sen ences, o pa ag aphs) side by side. E e y documen (legal ac ) was p ocessed o ge one column o each language. In o de o achie e his, bulle s and numbe ing, as well as nu- me ical pa o ables in documen s we e emo ed. Each language in he sou ce ile was p ocessed in o a sepa a e monolingual co pus and aligned wi h he co esponding co pus in he o he language. Table 1. size o he pa allel english-C oa ian co pus EnGLISH c oA IAn EnGLISH c oA IAn okens 1,151,297 1,059,406 Sen ences 72,045 75,638 Wo ds 951,156 855,560 Documen s 178 178 Apa om ex ac ing mos equen single-wo d and mul i-wo d e ms used as candida es o seman ic ames elemen s, he co pus can also be used o he ex ac ion o e ms, de ini ions and examples o he linguis ic con ex ha will be en e ed in o he e minological da abase Ai F ame. he co pus will also se e o he analysis o he syn ac ic le el o e ms, i.e. hei col- loca ions, ph aseological ela ions and p edica e-a gumen cons uc ions. 2.2. Te m ex ac ion and e m e i ica ion Au oma ic e m ex ac ion was conduc ed o each language wi h he op ion o ex ac ing a lis o 1000 single-wo d and mul i-wo d keywo ds. he EU -Lex English 2/2016 co pus was used as a e e ence co pus o ex ac ing English single-wo d e m candida es, while he English Web 2013 was used as a e e ence co pus o ex ac ing mul i-wo d e m can- dida es. Simila op ions we e possible o e m ex ac ion in c oa ian. he EU -Lex c oa ian 2/2016 co pus se ed as a e e ence co pus o ex ac ing mo e ele an single-wo d e m candida es, while his op ion was no possible o ex ac ing mul i- e m keywo ds. he e o e, he c o- a ian Web h Wac 2.2. co pus was used ins ead. In o de o es he eliabili y o he me hod, he i s 100 e m candida es aken om he English lis o mul i-wo d e ms we e manually alida ed. Ha ing checked hei conco dances, ou o 100 e m candida es, 70 we e ecognized as alid e ms by a e minologis wi h ample expe ience in 7 9Te minologija | 2019 | 26 a ia ion e minology. he lis o English keywo ds (o single-wo d e ms) consis ed la gely ou o a ia ion abb e ia ions and a ew key a ia ion e ms. Manual analysis and e m e i ica ion o bo h English and c oa ian candida e lis s o ex ac ed e ms was hen conduc ed. he e i ica ion p ocess consis ed o wo phases: 1. emo ing all abb e ia ions and ob ious gene al language wo ds and colloca ions; and 2. analysing he conco d- ances o he emaining e ms in o de o e i y whe he hey we e indeed a ia ion e ms. Abb e ia ions a e commonly used in a ia ion as e ms, hus no including hem in a dic iona y o glossa y o a ia ion would be unde s ood as a poo p esen a ion o a ia ion e minology in use. How- e e , since we wan ed o a i e a a lis o e ms o mul i-wo d cons uc- ions ha could be used in naming he elemen s o seman ic ames in a ia ion, abb e ia ions as lexical subs i u es o ull e ms we e no ea ed in he same unc ion he e. Apa om his, he e was no eason o keep- ing abb e ia ions in he c oa ian e m candida es lis s because hey we e used in he ex s as o eign language elemen s. 2.3. Ex ac ion and analysis o e bs he lis s o English and c oa ian e b we e ex ac ed au oma ically using he Keywo ds op ion in Ske ch Engine. Ve bs be, ha e and do ha e no been aken in o accoun in he English co pus, while bi i “beˮ, ima i “ha eˮ, moći “canˮ and mo a i “mus ˮ ha e been omi ed om he analy- sis in he c oa ian co pus. he i s 100 e bs om each lis a e hen analysed by manual alida ion o hei wo d ske ches. colloca ions deemed as mul i-wo d e ms o colloca ions o ela i e signi icance o a ia ion e minology based on hei equency a e no ed down in an Excel shee nex o he analysed e b. 3. ESUL S AnD DIScUSSIon Manual e m e i ica ion o single-wo d and mul i-wo d English and c oa ian e m candida es yielded su p ising esul s conce ning p ecision and ecall o he au oma ic ex ac ion p ocess in Ske ch Engine. Fo in- s ance, a e he i s phase o e m e i ica ion, ou o 1000 e ms, only 400 single-wo d c oa ian e ms emained ha we e no abb e ia ions o gene al language wo ds. he inal lis s o c oa ian e ms consis o 199 single-wo d and 455 mul i-wo d c oa ian e ms. he e a e somewha mo e English mul i-wo d e ms, bu no as many as one would expec . 664 mul i-wo d English e ms emained a e he analysis, bu only 148 8 0 Ana Os oški Anić Ex ac ing Lexical Uni s o Iden i ying Maja Lonča Specialized Seman ic F ames Ma ina Pa ić single-wo d English e ms, which makes o a signi ican ly poo esul o only 14.8% alid e ms om he ini ial lis o mul i-wo d English e ms. As expec ed, he second phase o e m e i ica ion p esen ed se e al issues ega ding he e mhood o he candida es. I bo h he noun and he adjec i e e e ing o a ce ain concep appea ed on he lis , he noun was kep as a mo e p o o ypical e m ealiza ion.7 his was mo e he case wi h he c oa ian e ms since in English a noun and an adjec i e o en ha e he same mo phological o m, as in he ollowing examples: kabina, kabinski “cabinˮ, na igacija, na igacijski “na iga ion, na iga ionalˮ, helikop- e , helikop e ski “helicop e ˮ, iden i ikacija, iden i ikacijski “iden i ica ionˮ. Simila ly, i bo h an adjec i e and an ad e b appea ed in he lis s, he adjec i e o m was kep : ins umen alni, ins umen alno “ins umen ˮ, while in he case o he e b/adjec i e di e ence, he e b was no mally kep on he lis : ce i ici a i, ce i ici an “ce i y, ce i iedˮ. When wo nouns e e ing o he same concep appea ed, bo h we e kep i one o hem was e e ing o he ac ion, and he o he o he esul , as in he example o iden i ikacija and iden i ici anje “iden i ica ionˮ. I seems ha no such di e ence exis s in he English coun e pa s. he English e m candida es lis s included a lo o in o ma ion ha needed o be il e ed, such as he names o coun ies, pe sonal names (e.g. juncke ), wo ds in o he o eign languages, and e en some unc ion wo ds (e.g. he p onoun his). he algo i hm o he ex ac ion o keywo ds and mul i-wo d e ms appa en ly ecognizes o eign elemen s, i.e. non-English wo ds in he co pus as e ms, oo. his could ha e been a oided had he ph ases (in all EU languages) ha usually appea a he end o each docu- men been e ased p io o adding he documen s o he co pus, bu i was decided ha i was easie o igno e hem in he candida es’ lis s a e wa ds. he second phase e m e i ica ion ega ding he English e m candi- da es showed he inabili y o dis inguish wo-pa om h ee-pa mul i- wo d cons uc ions as he g ea es pi all o he applied algo i hm, which is a common e o in many s a is ical e m ex ac ion ools. Bound ai and simula ed engine a e hus pa s o la ge , co ec mul i-wo d e ms bound ai ca go and simula ed engine ailu e (bo h e sions appea ed on he lis ), as well as ins umen g ound and c ew compa men (ins ead o 7 his kind o easoning is mo e in line wi h adi ional e minology wo k, in which nouns had p io i y in e minological dic iona ies o e co esponding adjec i es and e bs e e ing o he same concep (Sage 1990: 58). Howe e , bo h would be included as lexical uni s connec ed wi h he seman ic ame hey in oke. 8 1Te minologija | 2019 | 26 ins umen g ound ime and ligh c ew compa men ). Some mul i-wo d e m candida es o mally appea as p ope e ms and i akes domain speci ic knowledge o label hem as inco ec o alse posi i e esul s, e.g.: isual ligh > isual ligh ules, p o iding ai a ic > p o iding ai a ic se ices, in es iga ion au ho i y > sa e y in es iga ion au ho i y, maximum ake-o > maximum ake-o mass, seconda y su eillance > seconda y su eillance ada . Sho e e sions a e o en simply synonyms, like con- olle licence o ai a ic con olle licence. Ano he g oup o alse posi i e mul i-wo d e ms consis s o mul i-wo d exp essions ha could be labelled as belonging o gene al language: el- e an ai c a , indi idual ai c a , speci ic ai c a , la ge ai c a , p esen ai c a , heigh o cloud base, ield o ci il a ia ion, en i e lying ime o speci ic aining. Al hough hey a e mo pho-syn ac ically alid o ms, hey e e o no pa icula concep and he e o e canno be conside ed e ms. Ha ing analysed all lis s, we a i ed a a e y small numbe o single- wo d e ms, 199 c oa ian and 148 English e ms. In o de o e i y whe h- e he e ms ha could be conside ed o deno e “key” concep s o he ield had been ex ac ed, he wo d ske ches o he i s 10 e ms in each language we e analysed. Wo d ske ches a e au oma ic, co pus-de i ed summa ies o wo ds’ g amma ical and colloca ional pa e ns, and hey se e as an excel- len sou ce o gene a ing a la ge numbe o e m candida es in hei syn ac ical con ex (Kilga i e al. 2014). Apa om ypical adjec i e/ noun (we -leased ai c a ) o noun/noun (ai c a noise, ai c a posi ion) colloca ions, e bal colloca ions (ope a e an ai c a ) and p eposi ional ph ases (ai c a en e ing in o) gi e in aluable e minological in o ma ion mos use ul o e minologis s and ansla o s. he mos ele an ela ion o his pu pose was he coo dina ion ela ion be ween he a ge wo d and he ela ed wo ds. As can be seen in Figu e 1, he wo ds ela ed o ai c a a e occupan , e ain, ehicle, p oduc , sys ems, con en , loss, uni , engine, equipmen and ype. We compa ed he lis o he coo dina ed e ms wi h he lis o he single-wo d e ms and ma ked hose ha appea ed in bo h. he second s ep in e i ying he esul s was compa ing he mos equen colloca ions o he e ms in wo d ske ches wi h hose om he lis o mul i-wo d e ms. When he colloca ions om he wo d ske ch (e.g. ai c a iden i ica ion o ai c a ope a o in Figu e 1) ma ched hose in he lis o e ms ex ac ed om he co pus, hey we e also ma ked as lexical uni s o be included in he desc ip ion o a seman ic ame.