Show, Don't Tell : Visualising Finnish Word Formation in a Browser-Based Reading Assistant

This is a sel -a chi ed e sion o an o iginal a icle. This e sion may di e om he o iginal in pagina ion and ypog aphic de ails. Au ho (s): Ti le: Yea : Ve sion: Copy igh : Righ s: Righ s u l: Please ci e he o iginal e sion: CC BY 4.0 h ps://c ea i ecommons.o g/licenses/by/4.0/ Show, Don' Tell : Visualising Finnish Wo d Fo ma ion in a B owse -Based Reading Assis an © 2020 ACL Published e sion Robe son, F ankie Robe son, F. (2020). Show, Don' Tell : Visualising Finnish Wo d Fo ma ion in a B owse -Based Reading Assis an . In D. Al e , E. Volodina, I. Pilan, H. Lange, & L. B. Bo in (Eds.), NLP4CALL 2020 : P oceedings o he 9 h Wo kshop on NLP o Compu e Assis ed Language Lea ning (pp. 37-45). LiU Elec onic P ess. Linköping elec onic con e ence p oceedings, 175. h ps://www.aclweb.o g/an hology/2020.nlp4call-1.4 2020 Show, Don’ Tell: Visualising Finnish Wo d Fo ma ion in a B owse -Based Reading Assis an F ankie Robe son Uni e si y o Jy ¨ askyl¨ a [email p o ec ed] Abs ac This pape p esen s he NiinMik¨ aOli?! ead- ing assis an o Finnish. The ocus is upon he simpli ied p esen a ion and isualisa ion o a wide ange o wo d-le el linguis ic phenom- ena o he Finnish language in a uni ied o m so as o bene i language lea ne s. The sys em is a ailable as a b owse ex ension, in ended o be used in-con ex , wi h au hen ic ex s, in o de o encou age ee eading in language lea ne s. 1 In oduc ion This pape p esen s an in elligen eading assis an o Finnish. The sys em, NiinMik¨ aOli?! (English: TheWha Now?!), p esen s wo d and idiom de ini- ions in-con ex . The sys em can be used h ough a web in e ace ei he as a dic iona y o by manually en e ing o copying ex in o a ex ield, o ide- ally, as a b owse ex ension o assis wi h eading Finnish web pages. When used as a b owse ex- ension, NiinMik¨ aOli?! p esen s wo d de ini ions in a sideba . The e is inc easing in e es in con ex ualised lea ning o ocabula y (Godwin-Jones,2018), and NiinMik¨ aOli?! aims o acili a e his in he con- ex o web pages. NiinMik¨ aOli?! can be clas- si ied as an ATICALL (Au hen ic Tex In elligen Compu e Aided Language Lea ning) sys em, de- ined by Meu e s e al. (2010a) as so wa e which p oduces enhanced inpu based on eal ex s. The ocus o his pape is upon NiinMik¨ aOli?!’s simpli ied, uni ied iew o he Finnish language, which uses in o ma ion isualisa ion echniques o “show a he han ell” lea ne s abou mo pholog- ical and wo d o ma ion ea u es. A p inciple aim This wo k is licensed unde a C ea i e Commons A i- bu ion 4.0 In e na ional Licence. Licence de ails: h p: //c ea i ecommons.o g/licenses/by/4.0/. is o a oid p esen a ions which ell lea ne s abou wo d-le el g amma ical ea u es such as echnical linguis ic language including La ina e names o nominal cases, a he op ing o highligh hei su - ace o ms. The use in e ace isualises he con- nec ion be ween su ace o ms, analy ic o ms and de ini ions. Desc ibed i s is he cons uc ion o he base- line eading assis an sys em. The sys em is by and la ge simila o exis ing sys ems such as GLOSSER (Ne bonne e al.,1998) o he eading assis an ea u es o SMILLE (Zilio e al.,2017) o Re i a (Ka inskaia e al.,2017) – o e en e y widely used sys ems such as he Wo dNe -based al e na i e ansla ions shown when a single wo d is selec ed in Google T ansla e. No able as an imp o emen o e some o hose sys ems is Ni- inMik¨ aOli?!’s use o a ull scale Wo d Sense Dis- ambigua ion (WSD) sys em. The es o his pape desc ibes he mo i a ion behind and implemen a- ion o NiinMik¨ aOli?!’s isualisa ions and i s ex- haus i e ea men o complex lexical i ems such as de i ed wo ds, compounds and mul iwo d ex- p essions. 2 Baseline sys em A combined lexical esou ce o Finnish was c e- a ed by combining wo sou ces: FinnWo dNe (Lind´ en and Ca lson,2010) and Wik iona y. The Wik iona y de ini ions we e ex ac ed om pub- licly a ailable dumps, using a Py hon sc ip 1. The Py hon sc ip pa ses MediaWiki ma kup in o wo d senses using mwpa se omhell2. A leas one de ini ion was ex ac ed om 99.8% o a o al o 153 196 Wik iona y pages con- 1A ailable a h ps://gi hub.com/ ankie / wikipa se. 2h ps://gi hub.com/ea wig/ mwpa se omhell F ankie Robe son 2020. Show, don’ ell: Visualising Finnish wo d o ma ion in a b owse -based eading assis an . P oceedings o he 9 h Wo kshop on Na u al Language P ocessing o Compu e Assis ed Language Lea ning (NLP4CALL 2020). Linköping Elec onic Con e ence P oceedings 175: 37–45. 37 aining Finnish as a sec ion heading3. O hese, 90 653 a e lemmas a he han in lec ed o ms. Fo compa ison, FinnWo dNe con ains 139 871 head- wo ds, which a e mos ly lemmas bu include oc- casional idioma ic wo d o ms such as humalassa (li e ally “in hops”). FinnWo dNe is modelled a e Wo dNe , and as such has e y ine-g ained sense dis inc- ions. This esul s in po en ially o e whelming he lea ne wi h oo much in o ma ion. Fu he mo e, some Wik iona y senses a e likely o essen ially duplica e FinnWo dNe . Thus, simila de ini ions should be clus e ed oge he and only he bes de - ini ion displayed by de aul . The clus e ing and alignmen was c ea ed using a ini y p opaga ion (F ey and Dueck,2007). The dis ances g aph is cons uc ed by aking cosine dis ances be ween de ini ions, ep esen ed as ec- o s based on he English ex o hei de ini ions acco ding o he English sen ence simila i y model o Reime s and Gu e ych (2019). This model is based on p e ained English BERT models ine- uned on a seman ic simila i y ask. The p e ained be -la ge-nli-s sb-mean- okens model is used. Links be ween Wik iona y de ini- ions wi h dis inc e ymologies a e hen emo ed and ex a weigh is gi en o Wik iona y de ini ions so as o encou age hem o become exempla s o clus e s. The esul ing sys em ob ains adjus ed and index sco es o 0.48 on a gold s anda d o Wo dNe e bs g ouped by P opBank sense ob ained om P edica e Ma ix (de Lacalle e al., 2016), and a sco e o 0.52 on a manually c ea ed clus e ing o 128 Wik iona y and Wo dNe noun de ini ions. The sciki -lea n (Ped egosa e al., 2011) implemen a ion o a ini y p opaga ion is used. WSD is pe o med using UKB (Agi e e al., 2014). Since UKB is a g aph based WSD al- go i hm, i only ope a es on de ini ions om FinnWo dNe , which a e connec ed by he se- man ic links om P ince on Wo dNe . In o de o compa e Wo dNe de ini ions wi h Wik iona y de ini ions, he clus e ing is used. Clus e s a e hen anked using hei bes Wo dNe de ini ion as a ep esen a i e. Since Wik iona y de ini ions a e usually be e o lea ne s, hey a e pushed o he op o each clus e in he use in e ace. 3In a Wik iona y dump om 6/4/2019. 100101102103104105106107 Rank 0.0 0.2 0.4 0.6 0.8 1.0 P opo ion Remaining p opo ion Compounds pe oken Figu e 1: P opo ions ela ed o wo ds unknown o a simpli ied model o a language lea ne . The x-axis gi es he ank o he wo d ha he lea ne has lea ned all wo ds up o. Remaining p opo ion is he p opo ion o wo ds in un- ning ex unknown o he language lea ne . Compounds pe oken is he p opo ion o unknown wo ds which a e com- pounds. 3 Linguis ic a ionale Finnish is mo phologically ich4. Subs an i es a e declined o case and numbe and e bs a e conjuga ed o pe son, ense and oice. Finnish wo d o ma ion is also ich5. I includes a numbe o highly p oduc i e de i a ional mo phemes, in- cluding many de e bal mo phemes which is cha - ac e is ic o he language. Compounding also plays a majo ole in Finnish wo d o ma ion, wi h many o he compounds being seman ically ans- pa en . Finnish also has a numbe o encli ic pa icles such as he ques ion o ming “-ko/-k¨ o”. Finally, i has MWEs (Mul i-Wo d Exp essions) such as idioms. In Finnish hese may ake he o m o syn ac ic ames, ea ed he e as gapped MWEs e.g. “pi ¨ a¨ a-s a”, which could occu in a o m such as “pid¨ an oileip¨ akakus a” (English: I like sandwich cake) dis inguished om e.g. “pid¨ an oileip¨ akakun” (English: I keep sandwich cake). Why bo he going o he e o o making a comp ehensi e ea men o wo d o ma ion and complex wo d ypes? A e all, hese lexical i ems occu ela i ely in equen ly in unning ex and so i may seem like a poo alloca ion o e o o spend ime dealing wi h hem. One assump ion he e is ha hese elemen s become mo e impo - an a e he beginne s age o language lea n- ing. I we assume a e y simpli ied model o lexical acquisi ion whe e wo ds a e lea n in de- 4See o example Ka lsson (2015). 5See o example Hy ¨ a inen (2019). P oceedings o he 9 h Wo kshop on Na u al Language P ocessing o Compu e Assis ed Language Lea ning (NLP4CALL 2020) 38 scending o de o equency, we can analyse p op- e ies o wo ds ha he language lea ne does no know and he e o e may like o look up. Figu e 1 shows wo such p ope ies a ying as he numbe o wo ds he lea ne knows inc eases: he p opo - ion o all wo ds seen which a e unknown, and he p opo ion o unknown wo ds which a e com- pounds. The da a is based on 1.5 billion okens o analysed Finnish ex om he Tu ku In e ne Pa sebank (Laippala and Gin e ,2014). Taken as a whole, he co pus is 9.8% compounds. Supposing ha an in e media e lea ne may know somewhe e be ween 1000 and 10 000 wo ds. A e lea ning 1000 wo ds, 24% o unknown wo ds would be compounds, and a e 10 000, i would be 42%. Thus qui e a la ge p opo ion o wo ds unknown o in e media e le el lea ne s a e compounds. I is assumed ha o he complex lexical i ems such as MWEs ollow a simila pa e n. He e we e e o any i em which can be gi en a de ini ion, includ- ing lemmas, indi idual mo phemes and MWEs as headwo ds. Admi ing hese i ems a e equen , he nex ques ion becomes, why is simple lemma ex- ac ion no su icien ? One a gumen agains pe o ming lemma ex ac ion and simply show- ing lemmas comes om he no icing hypo he- sis (Schmid ,1990) which s a es ha wi hou a - en ion o o m (Ligh bown and Spada,2013, pp. 168–175), second language lea ne s in pa ic- ula a e p one o no acqui ing ine-g ained g am- ma ical knowledge. Following his concep , sys- ems such as hose o Meu e s e al. (2010b) and Reynolds e al. (2014) we e c ea ed o au oma i- cally enhance inpu in web pages in o de o p o- mo e no icing o , o example, pa s o speech. Ni- inMik¨ aOli?! ollows a simila di ec ion, bu in- s ead ocusses on mo phemes, d awing a en ion o he connec ion and o e lap be ween analy ic and su ace o ms, so o p omo e lea ning o mo - phology, as well as a en ion o he o ma ion o he wo d i sel . Why analyse wo ds using only no malised seg- men s, a he han — as a lo o e e ence ma e- ial o he Finnish language does — using g am- ma ical desc ip ions, such as La ina e names o case endings. The eason o his is wo old. Fi s ly, as Bleyhl (2009) no es, ea men s o lan- guage which a e hea y on g amma ical analysis and he associa ed linguis ic e minology can be coun e -p oduc i e in language ins uc ion since hey d aw a en ion away om he comp ehen- sible inpu needed o ue language acquisi ion. This la ge amoun o ex a ma e ial can lead o educed con idence om lea ne s. Secondly, due o Finnish’s agglu ina i e mo phology, con as ed wi h a usional language like La in, i is simply no necessa y o add his ex a laye o analy i- cal language, since many Finnish in lec ional mo - phemes occu in he same o m o an easily ecog- nisable o m a all imes, and hey can hus be e- e ed o by hei no malised o m. Conside o example, he Finnish sys em o loca i e case end- ings. These ha e a ai ly good co espondence in e ms o unc ion wi h p eposi ions in English. Imagine i , when eaching English, e e y p eposi- ion was also gi en a name o desc ibe i so ha we would always e e o “ om” as “ he ela i e p eposi ion”. I is ha d o imagine ha a lea ne would be well se ed by his ex a indi ec ion! This p inciple is somewha lexible, howe e , and he names o he mos common case endings — pa i i e and geni i e — a e shown on he basis ha hei usage is mo e g amma ical. They a e mo e o en obliga ed by con ex a he han used wi h he in en ion o con eying ex a in o ma ion. Plu al is e e ed o by-name since i is likely o be amilia . 4 Implemen a ion The pipeline om unning ex o analy ical seg- men s, desc ibed in his sec ion, is shown in Fig- u e 2. 4.1 MWE lexicon & ex ac ion FinnMWE (Robe son,2020) is used as a lexicon o MWEs. In o de o ex ac MWEs om unning ex , o each MWE is indexed by all possible lem- mas o a key oken. In case he head is known, i is used as he key oken, o he wise he a es o- ken based on wo d eq (Spee e al.,2018) is used. MWEs a e hen ex ac ed om dependency ees ob ained using he Tu ku neu al dependency pa s- ing pipeline (Kane a e al.,2018). Fi s , all key ma ches a e ound simul aneously by looking up all lemmas in he dependency ee. These candi- da es a e il e ed by ying o ma ch each emain- ing oken in he MWE agains any neighbou , ex- ending he neighbou hood in he p ocess un il he whole MWE is ma ched o i is impossible o p o- ceed. When he MWE key is i s head, i s pa en is excluded om he neighbou hood o po en ial P oceedings o he 9 h Wo kshop on Na u al Language P ocessing o Compu e Assis ed Language Lea ning (NLP4CALL 2020) 39 Running ex Token Dependency ee MWE Headwo d Analy ical segmen Tokenisa ion Tu ku-neu al- pa se -pipeline Reuse ancho Reuse ancho Analy ical segmen a ion s ep Ex ac ion Reuse ancho Spli Fo ce-Align Analy ical segmen a ion s ep Fo ce-Align Figu e 2: Diag am showing p ocessing pipeline om su ace o ms o analy ical segmen s. Objec s indica ed in yellow a e linked wi hin he use in e ace du ing he ho e b ushing in e ac ion. Do ed lines indica e how spans co esponding o he in-node a e ound in he ou -node. ma ching okens. MWE okens wi hou a lemma ac as wildca ds, and can ma ch mul iple okens, bu hey mus be connec ed wi hin he dependency ee.6 4.2 Analy ical segmen a ion da a The app oach o analy ical segmen a ion pu - sued he e is o combine analyses om he Omo i mo phological analyse (Pi inen,2015) and in o ma ion om Wik iona y oge he o p oduce analy ical segmen a ions. Omo i p o- duces analyses in i s own o ma , which has some deg ee o compa ibili y wi h ags om he Uni e sal Dependencies (UD) p ojec (Pyysalo e al.,2015). As an example, kakus a may be analysed as [WORD ID=kakku] [UPOS=NOUN][NUM=SG][CASE=ELA]. Mo - phological ags a e mapped o analy ical mo - phemes so ha e.g. [CASE=ELA] is ou pu as -s a, while WORD ID is passed h ough. The o - de in which he ags appea is he same o de as he su ace mo phemes appea , meaning ou an- aly ical mo phemes a e in he same o de as he 6The MWE ex ac ion code is made a ailable a h ps: //gi hub.com/ ankie /lex ac su ace mo phemes. Wik iona y con ains a ious empla e ags which gi e in o ma ion abou wo d o ma ion. This da a is sc aped in o a da abase so ha each e ymology sec ion can gi e a de i a ion o i s headwo d. Templa e ags a e no malised in o ei- he in lec ions, de i a ions o compounds consis - ing o no malised segmen s. Fo compounds and mos de i a ions, no malised segmen s a e di ec ly a ailable as a gumen s o he empla e ag. How- e e , he agen noun o empla e ag, o ex- ample, mus be manually mapped o “-ja”. Finally, he o m o empla e ag makes use o g am- ma ical e ms such as ela i e, which a e mapped o no malised segmen s such as “-s a”. 4.3 Building segmen a ion de i a ions Complex wo ds may ha e se e al le els o com- pounding o wo d de i a ion and in lec ion. Thus, we may ha e o make use o se e al lookups o ully segmen a wo d o m. We also wan o make su e a comple ely segmen ed wo d o m can be as- socia ed wi h all lexical i ems ha make i up. We hus shi ou pe spec i e o hink o hese analyses as ules and he segmen e as a ule engine which applies hem o p oduce de i a ions subjec o con- s ain s. Each ule can ma ch any single segmen and p oduce many segmen s. The basic ule engine ope a es by ecu si ely applying ules. I keeps ack o he cu en on o he de i a ion ee. A each i e a ion, each node om he on is conside ed and one o mo e s eps consis ing o applying one o mo e ules a e aken o c ea e child nodes, c ea ing a new on . The e may be mul iple ules which can ma ch a segmen . In his case, all combina ions o ules ma ching each ma chable segmen a e applied. When ei he he e a e no mo e ules which ma ch, o he e is a ma ch which does no expand any segmen s, he node is ma ked as e minal. A simple app oach would be o allow all ules o apply a once. Howe e , Omo i analyses do no wo k e y well as ules as-is in ou case, o ex- ample o oileip¨ akakus a Omo i p oduces h ee di e en analyses o di e en le els o decom- pounding o oileip¨ akakku. I we we e o apply each o hese analyses as ules we would end up wi h 3 inal segmen a ions. Howe e , o ou pu - poses, hey should all be subsumed unde he same de i a ion. The e o e we ake he ollowing ap- p oach: P oceedings o he 9 h Wo kshop on Na u al Language P ocessing o Compu e Assis ed Language Lea ning (NLP4CALL 2020) 40 1. Fi s , apply Wik iona y based ules ecu - si ely. 2. Fe ch all Omo i ules esul ing om looking up he whole wo d o m 3. While he e a e Omo i ules le : (a) Remo e any Omo i ules al eady sub- sumed by he cu en de i a ion. (b) I any ules emain, apply he one p o- ducing he leas new segmen s and dis- ca d. (c) Apply Wik iona y based ules ecu - si ely. 4. Apply any e o i ing ules, which exis o deal wi h occasional cases o usional Finnish mo phology. Fo example oileip¨ akakus a (English: ou o / om sandwich cake) would p oduce he ollow- ing de i a ion: Example 1: oileip¨ akakus a → oileip¨ akakku s a Omo i: oileip¨ akakus a → oileip¨ a kakku s a Wik iona y: oileip¨ akakakku → oi leip¨ a kakku s a Wik iona y: oileip¨ a While oimakkaammin (English: mo e powe - ully) would p oduce he ollowing: Example 2: oimakkaammin → oimakkaas i mpi Wik iona y: oimakkaammin → oimakas s i mpi Wik iona y: oimakkaas i → oima kas s i mpi Wik iona y: oimakas → oida ma kas s i mpi Wik iona y: oima In his case a e o i ing ule mmin →s i mpi7can be applied: Example 3: oimakas s i mpi ← oimakas mmin Re o i : mmin ← oimakkaammin Connec o pa en 4.4 Cons ain s upon ules Applied as-is, his scheme will p oduce impossi- ble segmen a ions. Howe e , i we conside he POS (Pa -O -Speech) o each segmen , we can place cons ain s o amelio a e his. We use a simple se o POS ags based on Wo d- Ne : Ve b, Noun, Ad e b, Adjec i e & Unknown. The UD POS ags used by Omo i and he Wik- iona y POS headings a e mapped in o his com- mon scheme. The mapping is lossy, o example, UD adposi ions a e mapped on o he Ad e b POS. All closed classes, in e jec ions and a ixes a e mapped o Unknown. No e ha cons i uen wo ds o Finnish compounds can be in lec ed wo ds, and 7“-s i” is ad e b o ming mo pheme like “-ly”, while “- mpi” is a compa a i e o ming mo pheme like “-e ”. oimakkaammin o i m a k a s m m i n o i m a k a s s i m p i o i d a m a cos : 13 cos : N/Acos : 0 cos : 4 Figu e 3: Alignmen wi hin de i a ion ee o oimakkaammin. Da k yellow po ions deno e su ace spans, while each o he whole yellow po ions including da k and ligh deno e he whole logical span. The cos o each alignmen acco ding o Fo ce-Align is shown nex o he pa en segmen . The dashed lines indica e he alignmen is no p oduced by Fo ce-Align bu ins ead ob ained om he unde lying ule. In his case: he syn he ic ule -mmin →-s i -mpi. so he e in lec ed o ms a e ea ed as ha ing he POS o hei lemma. The pe missible compound POS pa e ns can hen be p oduced by a lis o p oduc ion ules, ob- ained by s udying Hy ¨ a inen (2019): Ve b →Noun Ve b (e.g., koe+len ¨ a¨ a) Ve b →Ad e b Ve b (e.g., edes+au aa) Noun →Noun Noun (e.g., oi+leip¨ a) Noun →Adjec i e Noun (e.g., puna+ iini) Adjec i e →Adjec i e Adjec i e (e.g., hy ¨ an+n¨ ak¨ oinen) We s a by ea ing he whole oken as ha ing Un- known POS, meaning we can ma ch any POS. A any ime a segmen is cons ained o ha ing one o a se o POS ags. Fo compounds o mo e han wo pa s, we can ob ain he possible POS pa e ns by expanding he p oduc ion ules gi en abo e. Re e ing back o Example 1, a Wik iona y ule allows he analysis o oi (Ve b) as oida (3 d pe s.), howe e oileip¨ akakku is known o be a Noun, meaning acco ding o he abo e ules, oi mus be ei he a Noun o an Adjec i e, meaning his ule canno be applied. 4.5 Ob aining alignmen s om de i a ions In o de o ind co espondences be ween indi- idual cha ac e s in analy ical segmen s, su ace o ms and headwo ds, we apply he Fo ce-Align p ocedu e a each s ep o he analy ical segmen- a ion de i a ion. Each child segmen is gi en a span in o he pa en segmen . These a e o de ed and non-o e lapping. Ma ches a e pe o med a - P oceedings o he 9 h Wo kshop on Na u al Language P ocessing o Compu e Assis ed Language Lea ning (NLP4CALL 2020) 41 Func ion FORCE-ALIGN(pa en s ing p, a ay o child s ings c1. . . cn) e u ns alignmen a C ea e a bounded p io i y queue pq wi h he lowes cos pa ial solu ion a i s on Add an emp y pa ial solu ion in o pq while he head o pq is no comple e do Pop pa ial solu ion j om on o pq /*Make a ma ch */ i j’s cu so in o chas no eached end and j’s cu so in o phas no eached end and p(j’s cu so in o p)=c(j’s cu so in o c) hen Add copy o jin o pq wi h i s cu so s in o pand cinc emen ed end /*Skip a pa en cha ac e */ i j’s cu so in o pis no a beginning o end hen Add copy o jin o pq wi h i s cu so in o pand i s pa en cha ac e s skipped inc emen ed end /*Skip he es o he cu en child segmen */ i j’s cu so in o pis no a beginning and j’s cu so in o chas no eached end hen Add copy o jin o pq wi h i s cu so in o cand i s child cha ac e s skipped inc emen ed end end a:= alignmen o med by solu ion a on o pq end Algo i hm 1: The Fo ce-Align p ocedu e o ind an alignmen be ween a pa en s ing and i s seg- men ed child en s ings. e no malisa ion. All s ings a e lowe cased and he on owels ¨ a, ¨ o and y a e mapped o he espec i e back owels a, o and u. We aim o minimise a cos de ined as he sum o he squa e o he numbe o pa en cha ac e s skipped and squa e o he numbe o child cha ac e s skipped. An example showing he ype o alignmen s p o- duced by Fo ce-Align applied o he de i a ion o oimakkaammin is shown in Figu e 3. Fo ce-Align is implemen ed as a dynamic p o- g amming s yle p ocedu e, gi en as pseudocode in Algo i hm 1. A each s ep, Fo ce-Align keeps ack o candida e solu ions in a p io i y queue, wi h he lowes cos pa ial solu ion always being a he on . The p io i y queue has bounded leng h o bound he unning ime — making he p oce- du e a o m o beam sea ch. Whene e a pa ial candida e solu ion is aken om he on o he queue, up o h ee new pa ial solu ions a e c e- a ed and added back o he p io i y queue: making a single cha ac e ma ch; skipping a single cha - ac e om he pa en s ing; and skipping he es o he cha ac e s in he cu en child segmen . The p ocedu e ends when he e is a comple e solu ion a he on o he queue. Each child segmen ’s ull span co e s he cha ac e s om he beginning o i s i s cha ac e ma ch o jus be o e he i s cha ac e ma ch o he nex child segmen , o un il he end o he pa en segmen in he case o he las child segmen . A span o a segmen on o any ances o seg- men can be ound by ollowing a simple ule a each s ep: shi he whole span igh wa ds by a enough o i any new segmen s o he le , and ex- end he igh edge o he end o he child span on o he pa en segmen while he cu en child segmen is he igh mos . As an example, con- side Figu e 3and he analy ical mo pheme -kas. I begins on he igh edge. We conside i s align- men on o oimakas and ind we mus shi i s le edge by i e cha ac e s o make space o oima. We eplace i s igh edge wi h ha o i s pa en , which does no change he span. -kas is s ill on he igh edge o oimakas when we conside he alignmen o oimakas and oimakkaammin. A his s ep, we do no ha e o shi he le edge since he e a e no new segmen s o he le . The igh edge is eplaced wi h he igh edge o he align- men o oimakas on o oimakkaammin, shi ing i igh by one cha ac e . The inal span con ains he cha ac e s ‘kkaa’ — which is he allomo ph co esponding o ‘kas’, as equi ed. When he use ho e s o e a segmen o a seg- men a ion in he use in e ace, he co esponding su ace o m o he same segmen should high- ligh . The highligh ing consis s o a s ong high- ligh o ha pa o he su ace o m which has o e lapping ex wi h he analy ic o m, and a weak highligh o ha pa which is g amma i- cally pa o he same mo pheme bu does no li e - ally ma ch. The s ong highligh is ound by ind- ing he longes ma ch be ween he child segmen and i s span wi hin he pa en segmen , while he weak highligh is made om any pa o he span which is le o e . Special conside a ion is gi en o wildca ds, such as -s a. In his case, ma ching is pe o med igh o le , and he wildca d is weakly ma ched agains ha which emains a e all o he analy ic segmen s a e aligned. P oceedings o he 9 h Wo kshop on Na u al Language P ocessing o Compu e Assis ed Language Lea ning (NLP4CALL 2020) 42 Figu e 4: A sc eensho showing he Finnish Wikipedia page Tu ku being ead using he b owse ex ension. 5 Visualising Finnish wo d o ma ion A sc eensho o he use in e ace is shown in Fig- u e 4. De ini ions a e g ouped by no malised seg- men a ion. Wi hin each no malised segmen a ion he e a e de ined headwo ds, each co esponding o one o mo e o he no malised segmen s. They a e o de ed in dec easing o de o co e age o he no malised segmen a ion, meaning hose de ini- ions which de ine he meaning o he su ace o m mos closely appea closes o he op. Wi hin each de ined headwo d appea s one o mo e clus e s o de ini ions, each wi h an exempla . To b ing a en ion back o su ace o ms om he no malised o ms, he in e ace highligh s he su ace o ms as he lea ne ho e s o e he seg- men ed o ms, as shown in Figu e 5. The in e ac- ion ecalls a one dimensional “ho e sc ub” ac- ion. Ini ially, he whole wo d o ph ase is ligh ly highligh ed. As he lea ne sc ubs o e analy ic mo phemes, he co esponding spans in he su - ace o m a e highligh ed. To show he connec ion be ween he no malised segmen a ion and i s de ini ions, pa s o he de- ined headwo ds a e highligh ed when no malised segmen s a e ho e ed o e , as shown in Fig- u es 4&5. The whole in e ac ion se es o link he di e en iews o su ace o m, analy i- cal o m and headwo ds.8 6 Conclusions and Fu u e Wo k This pape p esen ed NiinMik¨ aOli?! The sys- em s eamlines he expe ience o using e e ence ma e ial by p esen ing i in-con ex , emphasising he mos impo an pa s and p esen ing simpli ied g amma ical analyses which do no ely upon ech- 8The NiinMik¨ aOli?! b owse ex ension and websi e a e a ailable a h ps://niinmikaoli. i/, while he analy ical segmen a ion code is a ailable a h ps:// gi hub.com/ ankie /asa i. Figu e 5: A composi e sc eensho showing di e en s ages o he in e ac ion esul ing when a use b ushes o e segmen s in he ex analyse . nical linguis ic ja gon, ollowing he p inciple o “show, don’ ell”. Clea ly, he ques ion o whe he sys ems such as NiinMik¨ aOli?! uly help language lea ne s is a pe inen one. Quan i a i e use e alua ion o ali- da e exis ing ea u es and poin o new ones is hus an impo an piece o u u e wo k. NiinMik¨ aOli?! gi es de ini ions in English. Adding common L1 languages o Finnish lea n- e s such as Swedish, Russian o A abic, as well as Finnish i sel could be a use ul addi ion. The cu en analy ical segmen e is ule based, and hus canno handle ou o ocabula y wo ds. A machine lea ning app oach such as ha o Kann e al. (2016) could be combined wi h he da a de- eloped he e o add ess his. A u u e di ec ion o all eading assis an s is be e p edic ion o language lea ne needs, which would lead o a sys em which knows be o ehand which ypes o eading assis ance would be bes o o e ei he by explici ly eques ing in o ma ion om he lea ne , o implici ly using in o ma ion om p e ious in e ac ions wi h he so wa e. P oceedings o he 9 h Wo kshop on Na u al Language P ocessing o Compu e Assis ed Language Lea ning (NLP4CALL 2020) 43 Re e ences Eneko Agi e, Oie Lopez de Lacalle, and Ai o So oa. 2014. Random walks o knowledge-based wo d sense disambigua ion.Compu a ional Linguis ics, 40(1):57–84. We ne Bleyhl. 2009. The hidden pa adox o o eign language ins uc ion. o : Which a e he eal o eign language lea ning p ocesses.Inpu Ma e s in SLA. Cle edon: Mul ilingual Ma e s, pages 137–55. B endan J. F ey and Delbe Dueck. 2007. Clus e ing by passing messages be ween da a poin s.Science, 315(5814):972–976. Robe Godwin-Jones. 2018. Con ex ualized ocabu- la y lea ning.Language Lea ning & Technology, 22(3):1–19. I ma Hy ¨ a inen. 2019. Compounds and mul i-wo d ex- p essions in Finnish: Compounds and Mul i-Wo d Exp essions, pages 307–336. De G uy e . Jenna Kane a, Filip Gin e , Niko Miekka, Akseli Leino, and Tapio Salakoski. 2018. Tu ku neu al pa se pipeline: An end- o-end sys em o he conll 2018 sha ed ask. In P oceedings o he CoNLL 2018 Sha ed Task: Mul ilingual Pa sing om Raw Tex o Uni e sal Dependencies, pages 133–142. As- socia ion o Compu a ional Linguis ics. Ka ha ina Kann, Ryan Co e ell, and Hin ich Sch¨ u ze. 2016. Neu al mo phological analysis: Encoding- decoding canonical segmen s. In P oceedings o he 2016 Con e ence on Empi ical Me hods in Na - u al Language P ocessing, pages 961–967, Aus in, Texas. Associa ion o Compu a ional Linguis ics. F ed Ka lsson. 2015. Finnish: an essen ial g amma . Rou ledge. Anisia Ka inskaia, Ja ad Nou i, and Roman Yanga be . 2017. Re i a: a sys em o language lea ning and suppo ing endange ed languages. In P oceedings o he join wo kshop on NLP o Compu e Assis ed Language Lea ning and NLP o Language Acquisi- ion, pages 27–35, Go henbu g, Sweden. LiU Elec- onic P ess. Maddalen Lopez de Lacalle, Egoi z Lapa a, I zia Ald- abe, and Ge man Rigau. 2016. A mul ilingual p ed- ica e ma ix. In P oceedings o he Ten h In e - na ional Con e ence on Language Resou ces and E alua ion (LREC’16), pages 2662–2668, Po o oˇ z, Slo enia. Eu opean Language Resou ces Associa- ion (ELRA). Ve onika Laippala and Filip Gin e . 2014. Syn ac ic n-g am collec ion om a la ge-scale co pus o in e - ne innish. In Human Language Technologies-The Bal ic Pe spec i e: P oceedings o he Six h In e - na ional Con e ence Bal ic HLT, olume 268, page 184. P. M. Ligh bown and N. Spada. 2013. How languages a e lea ned, 3 d ed edi ion. Ox o d handbooks o language eache s. Ox o d Uni e si y P ess. K is e Lind´ en and Lau i Ca lson. 2010. Finnwo dne – innish wo dne by ansla ion.LexicoNo dica– No dic Jou nal o Lexicog aphy, 17:119–140. De ma Meu e s, Ramon Ziai, Luiz Ama al, Ad iane Boyd, Aleksanda Dimi o , Vanessa Me cal , and Niels O . 2010a. Enhancing au hen ic web pages o language lea ne s. In P oceedings o he NAACL HLT 2010 Fi h Wo kshop on Inno a i e Use o NLP o Building Educa ional Applica ions, pages 10–18. Associa ion o Compu a ional Linguis ics. De ma Meu e s, Ramon Ziai, Luiz Ama al, Ad iane Boyd, Aleksanda Dimi o , Vanessa Me cal , and Niels O . 2010b. Enhancing au hen ic web pages o language lea ne s. In P oceedings o he NAACL HLT 2010 Fi h Wo kshop on Inno a i e Use o NLP o Building Educa ional Applica ions, pages 10–18. Associa ion o Compu a ional Linguis ics. J. Ne bonne, D. Dok e , and P. Smi . 1998. Mo pho- logical p ocessing and compu e -assis ed language lea ning.Compu e Assis ed Language Lea ning, 11(5):543–559. Fabian Ped egosa, Ga¨ el Va oquaux, Alexand e G am- o , Vincen Michel, Be and Thi ion, Oli ie G isel, Ma hieu Blondel, Pe e P e enho e , Ron Weiss, Vincen Dubou g, e al. 2011. Sciki -lea n: Machine lea ning in py hon. Jou nal o machine lea ning esea ch, 12(Oc ):2825–2830. Tommi A Pi inen. 2015. De elopmen and use o com- pu a ional mo phology o innish in he open sou ce and open science e a: No es on expe iences wi h omo i de elopmen . SKY Jou nal o Linguis ics, 28:381–393. Sampo Pyysalo, Jenna Kane a, Anna Missil¨ a, Ve onika Laippala, and Filip Gin e . 2015. Uni e - sal dependencies o Finnish. In P oceedings o he 20 h No dic Con e ence o Compu a ional Linguis- ics (NODALIDA 2015), pages 163–172, Vilnius, Li huania. Link¨ oping Uni e si y Elec onic P ess, Sweden. Nils Reime s and I yna Gu e ych. 2019. Sen ence- BERT: Sen ence embeddings using Siamese BERT- ne wo ks. In P oceedings o he 2019 Con e ence on Empi ical Me hods in Na u al Language P ocess- ing and he 9 h In e na ional Join Con e ence on Na u al Language P ocessing (EMNLP-IJCNLP), pages 3982–3992, Hong Kong, China. Associa ion o Compu a ional Linguis ics. Robe Reynolds, Edua d Scha , and De ma Meu e s. 2014. A VIEW o Russian: Visual inpu enhance- men and adap i e eedback. In P oceedings o he hi d wo kshop on NLP o compu e -assis ed lan- guage lea ning, pages 98–112, Uppsala, Sweden. LiU Elec onic P ess. P oceedings o he 9 h Wo kshop on Na u al Language P ocessing o Compu e Assis ed Language Lea ning (NLP4CALL 2020) 44