Using WordNet for Building WordNets

c m p -l g /9806016 23 J u n 1998Using WordNet for Building WordNets Xavier Farreres, German Rigau, Horacio Rodríguez Departament de Llenguatges i Sistemes Informàtics.Uni versi tat Poli tècni ca de Catalunya. Barcelona. Spai n.{farreres, horaci o, g.ri gau}@lsi.upc.esAbstractTh i s paper summar ises a set of methodologies and techniques for the fast constructi on of multi li ngual WordNets.The Engl sh WordNet s used n th sapproach as a backbone for Catalan andSpan ish WordNets and as a lex ical knowledge resource for several subtasks.1 Motivation and Introduction One of the main issues in last years as regards NLP act i v i t i es i s the i ncreas i ngly fast development of generic language resources. A lot of such resources, nclud ng both software and l ingware i tems (lex i cons, lex i cal databases,grammars, corpora marked in several ways) have been made avai lable for research and i ndustri al appli cati ons.Special interest presents, for knowledge-based NLP tasks, the ava lab l ty of w de coverage ontologies. Most known ontologies (as GUM, CYC,ONTOS, MICROKOSMOS, EDR or WORDNET,see [Gomez 98] for an extensi ve survey) di ffer i n great extent on several characteri sti cs (e.g. broad coverage vs. domai n speci fi c, lexi cally ori ented vs. conceptually ori ented, granulari ty, ki nd of i nformati on placed i n nodes, ki nd of relati ons,way of bui ldi ng, etc.). It i s clear, however, that for a wi de range of appli cati ons, WordNet (WN)[Miller 90] as become a de-facto standard.The success of WordNet has determi ned the emergence of several projects that a i m the construction of WordNets for other languages than English (e.g., [Hamp & Feldweg 97], [Artale et al.97]) or to develop multi li ngual WordNets (the most important project i n th i s l i ne is EuroWordNet (EWN)1).1http://www.let.uva.nl/~ewn/ The aim of EWN project i s to bui ld a multi li ngual database wi th WordNets for several european languages (i n the fi rst phase, Dutch,Italian and Spanish in addition to English).The constructi on of a WN for a language Lg(LgWN) can be tackledi n d i fferent ways accord ng to the lex cal sources ava lable. Of course the manual construction can be undertaken qu te stra ghtforwardly and leads to the best results in terms of accuracy, but has the important drawback of i ts cost. So, other approaches have been carr i ed out tak i ng prof i t of ava i lable resources i n fully automati c or semi -automati c ways.Whi ch are these lexi cal resources? Basi cally four kinds of resources have been used: 1) English WN (EnWN), as an i ni ti al skeleton for tryi ng toattach the words of Lg to i t, 2) already existi ng taxonomies of Lg (both at word and at sense level),3) bilingual (English and Lg) and 4) monolingual (Lg) dictionaries. All the approaches using EnWN as skeleton are based on the assumption of a closeconceptual si mi lari ty between Engli sh and Lg, i n such a way that most of the structure (relations) in EnWN could be maintained for LgWN.In the case of bi li ngual di cti onari es the usual approach is to try to link the English counterpart of entries to synsets in EnWN and to assume that the entry can be linked to the same synset.Monol i ngual d i ct i onar i es have been used basi cally as a source for extracti ng taxonomi c (hypernym) links between words (or senses [Bruce & Guthri e 92], [Ri gau et al. 97]) and i n lower extent for extract ng other k nds of semant c relations [Richardson 97] (e.g. meronymic links).Once a taxonomy of Lg (already exi sti ng or bui lt from a monoli ngual MRD) i s avai lable, the task can consi st of 1) enri chi ng the taxonomi c structure wi th other semanti c li nks (manually or automat i cally), as is the case of bu i ld ing individual WNs, or 2) merging this structure with other already exi sti ng ontologi es (as EnWN or EWN).Th i s paper presents our approach to the constructi on of WNs for two languages, Spani sh and Catalan, and li nki ng the fi rst one to EWN.We have developed a methodology that uses as core source EnWN 2. The methodology implies 1)2We have used WordNet 1.5.The use of EnWN for guiding the selection of the basic concepts of our WNs, 2) the use of EnWN as skeleton for linking Spanish and Catalan words to English synsets using bilingual dictionaries, 3) theuse of EnWN, together w i th b i l i ngual and monol i ngual d i ct i onar i es for allow i ng theconstruction of taxonomies (at sense level) of our languages and 4) the use of EnWN together withalready bui lt fragments of SpWN and CtWN for merging and incorporating these taxonomies to our WNs.In sect on 2 an overall descr pt on of our approach is given. Sections 3, 4 and 5 focus on the procedures for extracti ng connecti ons between words/senses/synsets. Secti on 3 i s devoted to procedures based on the use of bilinguals, section 4on the constructi on of taxonomi es and secti on 5deals w th the merg ng method. In all thesesecti ons we wi ll enphasi ze the role played by EnWN as Knowledge Source. Secti on 6, fi nally,presents some conclusions of our work.2 Our way of building WordNets As we have pointed out in the introduction, our ai m has been to desi gn a methodology (and a software env i ronment support i ng i t) forfaci li tati ng the task of bui ldi ng WNs from oursources. As we are nvolved n EWN project (covering the Spanish part), the methodology has been defi ned to be compati ble whi ch the generalapproach, guidelines and landmarks of the wholeproject but also to allow a parallel development ofthe CtWN.The general approach for bui ldi ng EWN i s described in [Vossen et al. 97]. Roughly speaking,the approach follows a top-down strategy tryi ng to assure a hi gh level of overlappi ng between languages, at least i n the hi ghest levels of the h erarchy, but reflect ng the language-spec f c lexi cali zati ons and provi di ng the maxi mum of freedom and flex i b i l i ty for bu i ld i ng the nd v dual WordNets. Bas cally t cons sts of three major steps: 1) Construct ion of core-WordNets for a set of common base concepts (around 800 nouns and 200 verbs), 2) enrichment of these sets prov i d i ng relat i onal l inks and incorporating their direct semantic contexts and 3)top-down extension of these core-WordNets.In our case two different approaches have been followed for dealing with nouns and verbs 3.3Although other categories can be included in EWN (and cross-category relati ons an be establi shed) only nouns and verbs have been i ntroduced unti l now i n our WordNets except for demostration purposes.In the case of verbs most of the work has been performed manually. The ma i n source of nformat on has been the P rap des database [Castellón et al. 97] that consi sts of 3,600 Engli sh verbs forms organi zed around Levi n's Semanti c Classes connected to WN1.5 senses. The database contai ns the theta-Gri ds speci fi cati ons for each verb (i ts semanti c structure i n terms of cases or themat i c roles), translat i on to Span ish andCatalan forms 4 and di athesi s i nformati on. The connecti ons extracted from thi s database were cross-validated with the information provided by bi li ngual di cti onari es i n order to i mprove thei r accuracy.In the case of nouns we have followed EWN strategy in the next way:1) The two h i ghest levels of EnWN (top concepts and di rect hyponyms) were manually translated i nto Spani sh (i ncludi ng vari ants). The results were f i ltered dropp i ng out words appeari ng less than fi ve ti mes as genus terms i nour monolingual dictionary [DGILE 87] or occurringless than 50 times in DGILE definition corpus 5 andless than 100 times in LEXESP corpus6.Thi s i ni ti al set (Spani sh core concepts, 361synsets) was then compared with base concept sets of other s tes of EWN (roughly the un on of i ntersect i on pa irs between languages was considered as the common base concepts set). The missing concepts in Spanish were manually added and vertically bottom up extended leading to the common Base Concept set (around 800 synsets).Catalan Base Concepts set was then built to cover the Spanish Base Concepts set.2) The enri chment of the BC set has been performed in two steps. First, using bilinguals as main lexical source, and then using other sources (ma i nly taxonom i es). These processes are described below.3 Using English WordNet with bilingualsWhen trying to build a lexical taxonomy from scratch, we can take profit of a preexisting lexical taxonomy, EnWN in our case, assuming it is well formed, as a skeleton of a taxonomy where we will f i ll i n the lex i cal data. Th is ensures several advantages: i t speeds up the constructi on of a large lexi con, as the only problem left i s the4Spani sh and Catalan are languages close enough for allowing a simultaneous development of lexical sources.5i.e. set of all definitions included in DGILE (1 million words)6balanced corpus of Spanish (5 Million words).deci si on where to attach the lexi cal data. Thereare also some problems: nobody ensures that thewellformedness of a lex i cal taxonomy for a language keeps true for another language, there must be semant i c closeness between both languages. We have therefore assumed that the structure of the WN taxonomy would suffice in the earlier stages of the construction of the our WNs.So, we need to choose synonyms in Spanish 7 for the Engli sh words present i n the ori gi nal synsets of WN. One way to fulfil our requirements is using bi li ngual di cti onari es (see [Kni ght & Luk 94],[Okumura & Hovy 94]). But we have to perform a sense disambiguation task in order to know whichsense of both words (the Spanish and the English one) is being referred. In other words, we have to deci de, for whi ch sense of the Spani sh word and for wh i ch synset i n WordNet a relat ion of synonymy is being defined.There i s also another m i nor problem to overcome, the unification of the two directions ofthe bi li ngual di cti onary, whi ch in few cases are symmetri cal, to collect all translati ons together.It i s true that uni fyi ng both di recti ons of the bi li ngual di cti onary i mpli es loss of i nformati on potenti ally i mportant (e.g. the order i n whi ch translati ons are wri tten i s relevant). But the lack of systemati c work i n the constructi on of the bilinguals makes this information of very doubtful utility.Thus, we have processed the b i l ingualscreati ng what we have called the homogeneousb i l i ngual, wh i ch i s a b i l i ngual w ith both directions mixed. Then, for each Spanish word, we have collected all the words gi ven as correct translations. And this has been the source for our work of attachment of Spanish words to WordNet synsets.Hav i ng collected all the translat ions of a Spani sh word together, we have then classi fi ed the words in classes depending on their behaviour.They can be class i f i ed i n three d i mens i ons:polysemy, structural and conceptual.In the polysemy di mensi on, we classi fy the words in classes depending on the number and kind of translati ons. For example, all entri es that have only one translati on fall i n the same class when thi s translati on i s monosemous i n WN terms; all entr i es that have several translat i ons fall i n another class when these translat i ons are polysemous.7Although we ilustrate the methodology considering only Spani sh, we performed the whole process for both Catalan and Spanish (and we provide results for both).In the structural di mensi on, we classi fy the words i n classes dependi ng on the relati on that the translati ons owns i n WN. For example, all entri es whi ch have several translati ons, shari ng some of them a common synset in WN, fall in the same category; all entr i es i n wh ich one translat i on i s a d i rect hyponym of other translation fall in the same category, etc.In the conceptual di mensi on, we apply the conceptual di stance formula (whi ch i s explai ned i n secti on 4.2.1.) on elements of the entri es. For example, all entr i es w i th a low conceptual distance between synsets of their translations fall in the same class.Each of these classes defi nes a set of entri es with the same behaviour. A confidence score has been assigned to each class by means of a manual vali dati on of a si gni fi cant sample extracted from them. We deci ded to accept the classes wi th a preci ssi on of 85% or more as classes of words to include in the first version of SpWN.Bilinguals can be used a step further stating a supposition: when several methods give the same result for the same Spanish word, the confidence for thi s attachment i ncreases. We have carri ed out an experi ment checki ng the classes i n pai rs,evaluat i ng the prec i ss ion of the set of ntersect ons, and n all cases the prec ss on increased. We have removed the cases where the preci si on was over 85%, the threshold appli ed i nthe previous experiment. This caused an increment of 40% of the original set of attachments.Furthermore, it is clear that if we merge more bi li nguals, the homogeneous resulti ng wi ll be larger, and wi ll then generate larger classes. But,what is even more important, the classes are more precise because some bilinguals lack the inclusion of some translations for some words. Table 1 shows the current fi gures of both CtWN and SpWN followi ng thi s approach (see [Atseri as et al. 97]and [Benítez et al. 98] for further detai ls of the whole process and tools used).Nouns Words Synsets Connections Spanish 23,21718,57841,293Catalan 5,2314,7237,193Verbs Spanish 3,0873,2197,960Catalan 3,3373,2199,078Table 1: current volumes of CtWN and SpWN.The last point to address is the extension of the intersection method to larger number of classes. If wi th two classes the i ntersecti on i ncreased the conf idence an equ i valent increase whenntersect ng larger numbers of classes can be expected.As a matter of fact, the extens on of theintersection method would be nothing more than perform ng a mult var ant stat st cal analys s,where each of the classes would be a factor. The nteresti ng result of thi s multi vari ant analysi s would be a formula wh ch could be used to calculate the value of the conf i dence of anattachment, depending on the number of classes in which it occurs.4 Building Taxonomies using WordNet4.1 Exploiting taxonomies from MRDsA strai ghtforward way of obtai ni ng a LgWN can be performed acqui ri ng taxonomi c relati ons from conventional dictionaries following a purely bottom up strategy. That s, 1) pars ng eachdefinition for obtaining the genus, 2) performing a genus disambiguation procedure, and 3) building anatural classification of the concepts as a concepttaxonomy w i th several tops. Follow i ng th i s purely descr pt ve methodology, the semant c pri mi ti ves of the LgWN could be obtai ned by collecti ng those di cti onary senses appeari ng at the top of the complete taxonomies derived from the di cti onary. By characteri zi ng each of thesetops, the complete LgWN could be produced. For DGILE, the complete noun taxonomy was derived using the automatic method described by [Rigau et al. 97]8.However, several problems arise due to a) the source (i .e., ci rculari ty, errors, i nconsi stenci es,omi tted genus, etc.) and b) the li mi tati on of the genus sense di sambi guati on techni ques appli ed (i.e., [Bruce et al. 92] report 80% accuracy usi ng automat c techn ques, wh le [R gau et al. 97]report 83%). Furthermore, the top d ct onary senses do not usually represent the semant c subsets that the LgWN needs to characteri ze i n order to represent useful knowledge for NLP systems. In other words, there i s a mi smatch between the knowledge di rectly deri ved from an MRD and the knowledge needed by a LgWN.To i llustrate the problem we are faci ng, let us suppose we plan to place the FOOD concepts i n the LgWN. Ne ther collect ng the taxonom es derived from a top dictionary sense (or selecting a8This taxonomy contains 111,624 dictionary senses and has only 832 di cti onary senses whi ch are tops of the taxonomy (these top d i ct ionary senses have no hypernyms), and 89,458 leaves (wh ich have no hyponyms). That s, 21,334 def n t ons are placed between the top nodes and the leaves.subset of the top di cti onary senses of DGILE)closest to FOOD concepts (e.g., substanc ia -substance-), nor collecti ng those subtaxonomi es starti ng from closely related senses (e.g., bebi da -dri nkable li qui ds- and ali mento -food-) we are able to collect exactly the FOOD concepts present in the MRD. The first are too general (they would cover non-FOOD concepts) and the second are toospec i f ic (they would not cover all FOOD dictionary senses because FOODs are described in many ways).All these problems can be solved using a mixed methodology. That i s, by attachi ng selected top concepts (and i ts der i ved taxonom ies) toprescribed semantic primitives represented in the LgWN. Thus, f i rst, we prescr i be a m i n i mal ontology (represented by the semantic primitives of the LgWN) able to represent the whole lexicon deri ved from the MRD, and second, followi ng adescr i pt ive approach, we collect, for everysemant i c pr i m i t i ve placed i n the LgWN, i tssubtaxonom i es. F i nally, those subtaxonom i es selected for a semanti c pri mi ti ve are attached to the corresponding LgWN semantic category.We used as semant i c pr i mi t i ves the 24lex i cographer's f i les (or semant i c f i les) into whi ch the 60,557 noun synsets (87,641 nouns) of WN are classi fi ed 9. Thus, we consi dered the 24semantic tags of WN as the main LgWN semantic primitives to which all dictionary senses must be attached. In order to overcome the language gap we also used a b i l i ngual Span i sh/Engl i sh d i ct ionary.4.2 Attaching DGILE dictionary senses to semantic primitivesIn order to classi fy all nomi nal DGILE senses with respect to WordNet semantic files, we used a similar approach to that suggested by [Yarowsky 92]. Th i s task i s d i v i ded into three fully automati c consecuti ve subtasks. Fi rst, we tag a subset (due to the di fference i n si ze between the monoli ngual and the bi li ngual di cti onari es) of DGILE di cti onary senses by means of a process that uses the conceptual di stance formula (see 4.2.1); second, we collect sali ent words for each semanti c fi le; and thi rd, we enri ch each DGILE9One could use other semanti c classi fi cati ons, such as Roget's Thesaurus [Yarowsky 92], the LDOCE semantic or pracmatic codes [Slator 91] or even better, a Spanish semant i c class i f i cat i on such as the "D i cc i onar io Ideológi co de la Lengua Española J. Casares" (DILEC).Really, when usi ng thi s methodology a mi ni mal set of informed seeds are needed. These seeds can be collected from MRDs, thesauri or even by i ntrospecti on. (see [Yarowsky 95]).di cti onary sense wi th a semanti c tag collecti ng ev i dence from the sal i ent words prev i ously computed.4.2.1 Attaching WordNet synsets to DGILE headwords.For each DGILE def n t on, the conceptual di stance between headword and genus has been computed usi ng WN1.5 as a semanti c net. We obtained results only for those definitions having English translations (using a bilingual dictionary)for both headword and genus. By computi ng the conceptual di stance between two words (w1,w2)we are also selecting those concepts (c 1i ,c 2j ) which represent them and seem to be closer with respect to the semanti c net used. Conceptual di stance i s computed using formula (1).(1) dist (w 1,w 2)=minc 1i ∈w 1c 2i ∈w 21depth (c k )c k ∈path (c 1i ,c 2i )∑That i s, the conceptual di stance between two concepts depends on the length of the shortest path 10 that connects them and the speci fi ci ty of the concepts in the path.In this way, we obtained a preliminary version of 29,20511 di cti onary defi ni ti ons semanti cally labelled (that i s, wi th WN lexi cographer's fi les)wi th an accuracy of 64% (61% at a sense level).That i s, a corpus (collecti on of di cti onary senses)classified in 24 partitions (each one corresponding to a semantic category).4.2.2 Collecting the salient words for every semantic primitive.Thus, we can collect the salient words (that is,those representat i ve words for a part icular category) using a Mutual Information-like formula (2), where w means word and SC semantic class.(2) AR (w ,SC )=Pr(w |SC )log 2Pr(w |SC )Pr(w )Intu i t i vely, a sal ient word 12 appears s i gn i f i cantly more often in the context of a10We only consider hypo/hypermym relations.11Due to the different sizes of the dictionaries used we only compute the conceptual distance for 31% of the noun dictionary senses.12Instead of word lemmas, this study has been carried out using word forms because word forms rather than lemmas are representative of typical usages of the sublanguage used in dictionaries.semanti c category than at other poi nts i n thewhole corpus, and hence is a better than average i ndi cator for that semanti c category. The words selected are those most relevant to the semanti c category, where relevance i s def ined as theproduct of salience and local frequency. That is to say, i mportant words should be di sti ncti ve and frequent.We performed the training process considering only the content word forms from d ct onarydef n t ons 13 and we di scarded those sali entwords wi th a negati ve score. Thus, we deri ved a lexicon of 23,418 salient words (one word can be a salient word for many semantic categories).4.2.3 Enriching DGILE definitions with WordNet semantic i ng the sal i ent words per category (or semanti c class) gathered i n the previ ous step we labelled the DGILE di cti onary defi ni ti ons agai n.When any of the sali ent words appears i n a def i n i t i on, there i s ev idence that the wordbelongs to the category i ndi cated. If several of these words appear, the evidence grows. We add together the r we ghts, over all words n the defi ni ti on, and determi ne the category for whi ch the sum is greatest, using formula (3).(3)W (SC )=AR (w ,SC )w ∈definition∑Thus, we obta i ned a second semant i cally labelled version of DGILE. This version has 86,759labelled defi ni ti ons (coveri ng more than 93% of all noun definitions) with an accuracy rate of 80%(we have ga ned, s nce the prev ous labelled version, 62% coverage and 16% accuracy).Although we used the 24 lexi cographer's fi les of WordNet as semanti c pri mi ti ves, a more fi ne-grained classification could be made. For example,all FOOD synsets are classi fi ed under <food,nutrient> synset i n fi le 13. However, FOOD concepts are themselves class i f i ed into 11subclasses (i.e., <y o l k >, <g a s t r o n o m y >,<comestible, edible, eatable, ...>, etc.). Thus, i f the LgWN we are planni ng to bui ld needs to represent <beverage, drink, potable> separately from the concepts <comestible, edible, eatable,...> a fi ner set of semanti c pri mi ti ves should be chosen, for nstance, cons der ng each d rect hyponym of a synset belonging to a semantic file also as a new semantic primitive or even selecting13After discarding functional words.for each semanti c fi le the level of abstracti on we need.4.3 Selecting the main top beginners for a semantic primitive Thi s secti on i s devoted to the locati on of the mai n top di cti onary senses for a gi ven semanti c pr m t ve n order to correctly attach all tssubtaxonomies to the correct semantic primitive inthe LgWN.In order to illustrate this process we will locate the mai n top begi nners for the FOOD di cti onary senses. However, we must consider that many of these top beginners are structured. That is, some of them belong to taxonomi es deri ved from other ones, and then cannot be di rectly placed wi thi n the FOOD type. Thi s i s the case of vi no (wi ne),which is a zumo (juice). Both are top beginners for FOOD and one is a hyponym of the other.First, we collect all genus terms from the whole set of DGILE di cti onary senses labelled i n the prev i ous sect i on w ith the FOOD tag (2,614senses), producing a lexicon of 958 different genus terms (only 309, 32%, appear more than once in the FOOD subset of dictionary senses).As the automati c di cti onary sense labelli ng i s not free of errors (around 80% accuracy)14 we can discard some senses by using filtering criteria.• Filter 1 (F1) removes all FOOD genus terms not assigned to the FOOD semantic file during the mapping process between the bilingual dictionary and WN.• Fi lter 2 (F2) selects only those genus terms whi ch appear more ti mes as genus terms i n the FOOD category. That is, those genus terms which appear more frequently i n di cti onary defi ni ti ons belonging to other semantic tags are discarded.• F lter 3 (F3) d scards those genus terms which appear with a low frequency as genus terms i n the FOOD semant i c category. That i s,infrequent genus terms (given a certain threshold)are removed. Thus, F3>1 means that the fi lteri ng cr i ter i a have d iscarded those genus terms appear ng n the FOOD subset of d ct onary definitions less than twice.At the same level of genus frequency, fi lter 2(removing genus terms which are more frequent in other semanti c categori es) i s more accurate than fi lter 1 (removi ng all genus terms the translati on14Most of them are not really errors. For i nstance, all fishes must be ANIMALs, but some of them are edible (that i s, FOODs). Nevertheless, all fi shes labelled as FOOD have been considered mistakes.of which cannot be FOOD). For instance, no error appears when selecti ng those genus terms whi ch appear 10 or more times (F3) and are more frequentin that category than in any other (F2), discarding only 3% of correct genus terms (see [Rigau et al. 98]for complete figures).4.4 Automatically building large scale taxonomies from DGILE The automati c Genus Sense Di sambi guati on task i n DGILE has been performed followi ng[R i gau et al. 97]. Th i s method reports 83%accuracy when selecting the correct hypernym by combining eight different heuristics using several methods and types of knowledge (two of theheuristics use WN).Once the mai n top begi nners (relevant genus terms) of a semanti c category are selected and every d i ct i onary def i n i t ion has beendisambiguated, we collect all those pairs labelledwi th the semanti c category we are worki ng on havi ng one of the genus terms selected. Usi ng these pa rs we f nally bu ld up the complete taxonomy for a given semantic primitive. That is,i n order to bui ld the complete taxonomy for a semanti c pri mi ti ve we fi t the lower senses usi ng the second labelled lexicon and the genus selected from this labelled lexicon.Although, both f nal taxonom c structures produce more flat taxonomi es than i f the task i s done manually, a few arrangements could be done at the top level of the automati c taxonomi es.Studyi ng the mai n top begi nners we can easi ly di scover an i nternal structure between them (for FOOD, 18 or 48 depend i ng on the cr i ter i a selected).Perform ing the process for the whole dictionary we obtained for F2+(F3>9) a taxonomic structure of 35,099 defi ni ti ons and for F2+(F3>4)the si ze grows to 40,754. Testi ng the results on FOOD taxonomies we achived 99% accuracy with the first criterion and 96% with the second.5 Extending and Filling Gaps.Up to now we have described a methodology to connect words from a language to a WN skeleton,and another methodology to build taxonomies.The words f i nally connected i n the f i rst process, apart from the prec i ss ion threshold cri teri on, do not follow any other cri teri on: they are not the most i mportant, nei ther the topmost nor the lowermost concepts i n the hi erarchy; the connecti ons are scattered all over the skeleton.The final set of words connected to the skeleton is random, and we don't have any control over i t.。