Document Indexing with a Concept Hierarchy Índice de Documentos con una Jerarquía de Conceptos

Transcription

1 Document Indexng wth a Concept Herarchy Índce de Documentos con una Jerarquía de Conceptos Alexander Gelbuh, Grgor Sdorov and Adolfo Guzmán-Arenas Computacón y Sstemas Vol. 8 Núm. 4, pp , CIC-IPN, ISSN , Impreso en Méxco Natural Language Processng Laboratory, Center for Computng Research (CIC), Natonal Polytechnc Insttute (IPN), Av. Juan de Dos Bátz s/n, Esq. Mendzábal, Col. Zacatenco, CP 07738, DF, Méxco. E-mal: gelbuh@gelbuh.com, sdorov@cc.pn.mx, a.guzman@acm.org; Artcle receved on aprl 13, 2004; accepted on march 15, 2005 Abstract Gven a large herarchcal concept dctonary (thesaurus, or ontology), the tas of selecton of the concepts that descrbe the contents of a gven document s consdered. A statstcal method of document ndexng drven by such a dctonary s proposed. The method s nsensble to naccuraces n the dctonary, whch allow for sem-automatc translaton of the herarchy nto dfferent languages. The problem of handlng non-termnal and especally top-level nodes n the herarchy s dscussed. Common sense-complant methods of automatcally assgnng the weghts to the nodes and lns n the herarchy are presented. The applcaton of the method n the Classfer system s dscussed. Keywords: Document Characterzaton, Document Comparson, Ontology, Statstcal Methods. Resumen Se consdera la tarea de la seleccón de los conceptos que descrben el contendo de un documento dado. Los conceptos se elgen de un dcconaro erárquco grande (un tesauro, o ben una ontología). Se propone un método estadístco para crear un índce de los documentos, guado por tal dcconaro. El método es robusto en cuanto a los errores en el dcconaro, lo que permte traducr tal dcconaro semautomátcamente en varos lenguaes. Se dscute el problema del uso de los nodos no termnales y especalmente de los nodos de alto nvel en la erarquía. Se presentan los métodos para ponderacón automátca de los nodos y vínculos en la erarquía de la manera en que concde con los crteros del sentdo común. Se dscute la aplcacón del método en el sstema Classfer. Palabras Clave: Caracterzacón de Documentos, Comparacón de Documentos, Ontología, Métodos Estadístcos. 1 Introducton We consder the tas of ndexng a document wth concepts as mappng the document nto the concept dctonary, assgnng to each concept n the dctonary a value that reflects ts relevance for the gven document. Thus, the document s represented by a hstogram of ts topcs. Say, a newspaper artcle can be about ndustry (60%), transport (20%), scence (10%), etc. Note that these are concepts ncluded n the dctonary rather than the ey words drectly mentoned n the document; what s more, the document mght not contan the word transport at all, but nstead contan the words trans, ralways, etc. Such a representaton of documents s mportant for nformaton retreval (Charabart et al. 1997), document classfcaton, text mnng (Feldman and Dagan 1995), nvestgaton of document collectons (Lght 1997), text understandng, etc. In document retreval, the documents are scored by the correspondence of ther man topcs to the user s request. In text mnng, data mnng technques are appled to dscoverng trends and devatons of the topcs of dscusson n the newspapers. In text understandng, topc detecton allows selectng the language model (Seymore and Rosenfeld 1997). In document classfcaton and text segmentaton (Ponte and Croft 1997), topc detecton has been the obect of extensve research n recent years. A large core of research has been devoted to automatcally learnng the classfcaton rules usng statstcal and lngustc methods (Apté et al. 1994; Bharat and Henznger 1998; Cohen and Snger 1996), machne learnng methods (Koller and Saham 1997), neural networs, self-organzng maps (Hyötynem 1996) and connectonst models (Le et al. 1994). In the maorty of these studes, the tas of automatc constructon of the topc herarchy s consdered. In ths artcle, however, we consder the (non-weghted) topc herarchy to be pre-defned, and concentrate on ts usng and on assgnment of the weghts to the nodes and lns. The problems arsng n the complaton and use of a concept herarchy depend dramatcally on ts sze. In some applcatons, there s a small set of predefned topcs, and a typcal document s related to only one topc. For example, ths 281

2 Alexander Gelbuh, Grgor Sdorov and Adolfo Guzmán Arenas s the case for a governmental recepton offce where the complants t receves from the ctzens are to be classfed to send them to exactly one of the departments of polce, or health, or envronment, etc. However, n the case of open texts, such as Internet documents or newspaper artcles, the set of possble topcs s large and not so well defned, and the maorty of the documents are related to several or many topcs at the same tme. Ths leads to the necessty of some structurng of the set of topcs. The most natural structure for the concepts s a herarchy. For example, f a document s related to the narrow topcs electons, government, and party, then t can be classfed as a document on poltcs. Thus, though most of exstng dctonary-based systems use flat topc dctonares eyword groups wthout any herarchcal structure n ths paper we use a herarchcal dctonary and specfcally address the ssue of determnng the contrbuton of the top-level concepts. Such a problem does not exst n the flat document categorzaton dctonares. We consder the lst of topcs to be large but fxed,.e., pre-defned. Our ndexng algorthm does not obtan the topcs drectly from the document body; nstead, t relates the document wth one of the topcs lsted n the system dctonary. The result s, thus, the measure (say, n percents) of the correspondng of the document to each of the avalable topcs. Unle the tradtonal categorzaton approach, we consder fuzzy categorzaton, when a document can be ndexed wth many categores wth ther correspondng weghts. In Fgure 1, a screen shot of our program, CLASSIFIER, s shown wth a hstogram of the topcs of a Spansh document. Fg. 1. Topc hstogram for a document A problem arses of the optmal, or reasonable, degree of detal for such categorzaton. For example, when descrbng the Internet news for an average reader, the categores le anmals or ndustry are qute approprate, whle for the descrpton of artcles on zoology such a dctonary would gve a trval answer that all documents are about anmals. On the other hand, for average reader of Internet news t would not be approprate to categorze the documents by the topcs mammals, herptles, crustaceans, etc., snce such a descrpton s too detaled for such a user. In ths paper, we wll dscuss the structure of the topc dctonary, the choce and use of topc weghts, and some practcal aspects of complaton of the herarchcal concept dctonary. 282

3 Document Indexng wth a Concept Herarchy 2 Concept Herarchy In (Guzmán-Arenas 1997 and 1998), t was proposed to use a herarchcal dctonary for determnng the man themes of a document. The proposed algorthm s based on ths dea. Unle some other methods of ndexng (Nwa et al 1997), our algorthm does not obtan the canddate topcs drectly from the body of the document beng analyzed. Instead, t reles on a large pre-exstng dctonary of topcs organzed n a tree. Non-termnal nodes of ths tree represent maor topcs, such as poltcs or nature. The termnal nodes represent the narrowest topcs such as electons or crocodles. Termnal topcs are assocated wth so-called eyword groups. A eyword group s a lst of words or expressons related to the stuaton descrbed by the name of the topc. Such words and expressons are drectly used n the text. For example, the topc relgon can be assocated wth the words le church, prest, candle, Bble, pray, plgrm, etc. Note that these words are connected nether wth the headword relgon nor wth each other by any standard semantc relaton such as subtype, part, actant, etc. Ths maes complaton of such a dctonary easer than that of a real semantc networ dctonary. However, such a dctonary s not a plan varant of a semantc networ such as WordNet, snce some words are grouped together that have no mmedate semantc relatonshp. Thus, such a dctonary cannot be obtaned from a semantc networ by a trval transformaton. Fgure 2 shows another example of a dctonary entry. Techncally, our CLASSIFIER program manages contact word combnatons n the same way as sngle words. Though the concepts are organzed n a tree, a eyword can belong to several concepts. Ths can be due to ether homonymy of the word (Krowetz 1997): e.g., bll belongs to money, law, tools, brds, or due to ntersecton of the topcs: e.g., grl belongs to chldren and women. Fg. 2. Herarchcal dctonary used by the system 3 The Naïve Algorthm: No Weghts The algorthm of document ndexng wth the concept thesaurus conssts of two parts: ndvdual (leaf) topc detecton and propagaton of the topcs up the tree. The frst part of the algorthm s responsble for detecton termnal topcs,.e., for answerng, ndvdually for each termnal topc, the followng queston: To what degree ths document corresponds to the gven topc? In our current 283

4 Alexander Gelbuh, Grgor Sdorov and Adolfo Guzmán Arenas mplementaton, ths s done basng on a plan lst of words correspondng to the topc, see Fgure 3. However, n general, a topc can be assocated wth a procedure. For example, to detect that a document represents an applcaton form relevant to some department of a government offce, t may be necessary to analyze the format of the document. Fg. 3. Countng eywords n a document In our mplementaton, for each eyword group, the number of occurrences of the words correspondng to each (termnal) topc s determned. These numbers are normalzed wthn the document,.e., dvded by the number of words n the document. The accumulated number of occurrences s consdered the measure of the correspondence between the document and the topc. Note that the values for ths measure of relevance are not normalzed snce the topcs are not mutually exclusve. The second part of the algorthm s responsble for the propagaton of the found frequences up the tree. Wth ths, we can determne that, say, a document mentonng the termnal topcs mammals, herptles, crustaceans, s relevant for the nontermnal topc anmals, and also lvng thngs, and also nature. The dfference between ndexng wth a plan lst of termnal topcs and wth propagaton of the frequences to nontermnal topcs can be seen n fgure 1 and fgure 4. Actually, n fgure 1, only termnal concepts are shown, whle fgure 4 shows all concepts, ncludng non-termnal ones, for the same document. Though the man termnal topc for ths document s solders and mltary lfe, the man topc n the whole tree s INSTITUTIONS. 1 Propagaton of the frequences s crucal for the whole dea of our method. It s necessary to mae use of the nontermnal nodes of the herarchy and to generalze the contents of the document to a degree allowng for ts matchng wth the user s queres contanng more general words than the ones mentoned drectly n the document. However, t presents the problem of overgeneralzaton: appled n the naïve way descrbed here, t always assgns the greatest relevance to the toplevel concepts, so that any document s ndexed wth the concepts obect, acton, etc., as ts man topcs. As an example of ths problem, we can see that n fgure 4, the concept ANY TOPIC has unreasonably hgh ranng. Below we wll show how to cope wth ths problem. 1 In Fgure 4, the non-termnal topcs are shown n captal letters. 284

5 Document Indexng wth a Concept Herarchy Fg. 4. Non-termnal concepts n the ndex 4 Relevance and Dscrmnaton Weghts In the algorthm descrbed above, smple word lsts were used to count the frequences reflectng the relevance of the topc for a document. Thus, a word n the document could be counted ether as one occurrence of the correspondng topc or zero, whch s too rgd. To defne quanttatve measures of relevance of the words for the topcs and quanttatve measure of mportance of the nodes of the herarchy, some numerc weghts can be used by the algorthm. There are two nds of such weghts: the degree of the connecton between a eyword (or a subtopc) and the correspondng topc s assocated wth a ln n the herarchy, whle the measure of mportance of an ndvdual concept for the user s assocated wth the ndvdual node. The classfcaton algorthm taes nto account these weghts. Namely, for the accumulated relevance of the topcs, t multples the number of occurrences of a word (or subtopc) by the weght and then multples the result by the weght w 4.1 Relevance and Dscrmnaton Weghts of the topc tself. w of the ln between the word and the topc, The frst type of the weghts s assocated wth the lns between words and topcs or between the nodes n the tree. 2 For example, f the document mentons the word carburetor, s t about cars? And f t mentons the word wheel? Intutvely, the contrbuton of the word carburetor nto the topc cars s greater than that of the word wheel: f the document mentons carburetor, then t s almost surely s about cars, but f t mentons wheel, then t s possbly about cars but could be about clocs, songs, pottery, etc. Thus, the ln between wheel and cars s assgned a less weght than for carburetor and cars. The algorthm of classfcaton taes nto account these weghts when complng the accumulated relevance (frequency) of the topcs. 2 Actually, the former type s a nd of the latter snce the ndvdual words can be consdered as termnal tree nodes related to the correspondng topc. 285

6 Alexander Gelbuh, Grgor Sdorov and Adolfo Guzmán Arenas It can be shown that the weght w of such a ln (between a eyword and a topc or between a topc and ts parent topc n the tree) can defned as the mean relevance of the documents contanng ths word for the gven topc: r = D r D n n w, (1) By all the avalable documents D, where s the measure of relevance of the document to the topc, and s the number of occurrences of the word or topc n the document. Ths equaton can be used for automatc tranng of the dctonary gven a document collecton mared up wth the relevance of the topcs. Unfortunately, we are not aware of any relable algorthm of automatc detecton of the measure of the relevance of r r n an ndependent way. Thus, to apply ths equaton drectly or tranng the dctonary, such a measure s to be estmated manually by the expert. However, n practce such a wor s usually too expensve. As a practcal alternatve, t s often possble for the expert to estmate the weghts w ntutvely at the stage of preparaton of the dctonary. The choce of the weght s based on the frequency of appearance of the word n general documents from the control corpus of the texts on any topc; n our case such texts were the newspaper ssues. As another practcal approxmaton, for narrow enough themes we can assume the hypothess that the texts on ths topc never occur n the control corpus (newspaper mxture). Then, gven the fact that we have ncluded the word n the dctonary and thus there s at least one document relevant for the gven topc, we can smplfy the expresson for the weghts as follows: w = D 1 n n, (2) snce the numerator of the quotent n (1) n case of narrow topcs can be consdered to be 1. Not surprsngly, ths gves the weght of the word related to a specfc topc to be the less the more ts frequency; for example, the artcles a and the have a (nearly) zero weght for any topc, whle the word carburetor has a hgh weght n any topc n whch t s ncluded. Sometmes a rare enough word, say, a noun bll, n ts dfferent senses s related to dfferent topcs (money, law, brds, geography, tools). For a more accurate analyss, some nd of competton between senses of the word for a specfc occurrence n the document s ntroduced. For ths, the senses of the word are mared n the topc dctonares (as bll 1, bll 2, etc.), and the weghts of occurrences of such a word are to be normalzed by ts dfferent senses (though the occurrences of the same sense are ndependent n dfferent topcs), wth the wegh of an ndvdual sense n each document beng proportonal to the relevance of the document for the gven topc: w w ~ w r = 1 w, (3) where s the weght of the -th sense of the gven occurrence of the word n the gven document, s the weght of the ln between ths sense of the word and the topc, the summaton n the frst equaton s made by all the topcs, and n the second by the senses of the gven word. 3 Snce r n ts turn depends on w, to avod teratve procedure, n practce we calculate r based on equal weghts w. w 3 However, ths technque s not very mportant for most cases, snce usually t does not change the order of the topcs for a document, but only maes the dfference between dfferent topcs more sgnfcant. Snce n many cases t requres manual marng up the senses (for the words for whch such nformaton s absent n the dctonary), we use t rarely n our program; for the maorty of the words we do not dstngush senses. 286

7 Document Indexng wth a Concept Herarchy Note that the weghts of the lns are a natural part of the concept herarchy tself. Ths component mparts a quanttatve character to the herarchy. However, here we dscuss the ssue of calculatng the weghts because of two reasons. Frst, there are avalable concept herarches, such as Roget thesaurus, WordNet (1998), Factotum SemNet (Cassdy 2000), etc. However, these dctonares do not nclude any quanttatve nformaton. Second, the relevance weghts, generally speang, depend on the tranng set the collecton of general texts. For example, for applcaton of our ndexng method to a collecton of techncal artcles, the ln between wheel and song would have much less weght than for ts applcaton to the general Internet documents. Ths effect wll be descrbed n more detal n the next secton. 4.2 Dscrmnaton Weghts The classfcaton algorthm descrbed above s good for answerng the queston s ths document about anmals? but not the queston what about s ths document?. Really, as we have mentoned, wth such an approach taen lterally, the answer wll be all the documents are about obects and actons, the top nodes of the herarchy. However, a reasonable answer s usually that a document s about crustaceans, or anmals, or lvng thngs, or nature, dependng on the user s needs and level,.e., on the degree of detals to whch the user s nterested n the area. Thus, we suggest that the answer to the queston what about s ths document? depends on the user. For example, f the document mentons lobsters, shrmps, crabs, and barnacles, then for a bologst the answer crustaceans would be the best, whle for a schoolchld the answer bology s better, and for an average newspaper reader, the answer nature. How can we guess ths wthout havng to explctly as the reader? Asng the reader about the desred detal degree s not a soluton because, frst, he or she wll probably even not understand the queston, and, second, t s not possble for the reader to quanttatvely specfy the mportance of hundreds of topcs n the herarchy. Thus, an automatc way of assgnng the mportance weghts s necessary. Our hypothess s that the unverse of the reader s the base of the documents to whch he or she apples the search or classfcaton. In other words, we assume that the reader s a specalst n the contents of the current database beng ndexed. Thus, the weghts of the relevance of topcs n our system depend on the current database. The man requrement to these weghts s ther dscrmnaton power: an mportant topc should correspond to a (consderable) subset of documents, whle the topcs that correspond to nearly all documents n the data base are probably useless, as well as too narrow topcs that correspond to few documents n the base. Thus, the weght w of a tree node can be estmated as the varaton of the relevance such a dscrmnaton power s to smply measure t as the dsperson: r r the topc over the documents of the database. A smple way to calculate ( ) w = r M M = D D r D where M s the average value of over the current database D, and s determned by the former algorthm, wthout w 4 tang nto account the value of. Wth ths approach, for, say, a bologcal database, the weght of the topcs le anmals, lvng thngs, and nature s low because all the documents equally menton these topcs. On the other hand, for newspaper mxture ther weght s hgh. 2 r (4) 5 Applcatons Wth the approach descrbed above, we have mplemented n the system CLASSIFIER several useful functons. The system can ndex the document wth ts man topcs, wth or wthout the propagaton of the frequences to the nontermnal nodes, see fgure 1 and fgure 4. The system allows vewng the documents by topcs, answerng the queston: for a selected topc, what documents are the most relevant? Ths corresponds to the tas of nformaton retreval, see fgure 5. 4 In a more precse manner, the nformaton theory can be appled to the calculaton of the weghts; we wll not dscuss here ths dea. Yet another approach to the propagaton of the frequences to the non-termnal nodes s dscussed n (Gelbuh et al. 1999b). 287

8 Alexander Gelbuh, Grgor Sdorov and Adolfo Guzmán Arenas Fg. 5. Documents ordered by relevance for a specfc topc An nterestng applcaton of the method s classfcaton of the documents by smlarty to a gven document, see fgure 6. The comparson s made based on the relevance of concepts from the dctonary for the two documents,.e., on ther representaton wth the topc hstograms, see, for example, (Gelbuh et al. 1999a). The herarchcal structure of the dctonary allows for tang nto account the user s nterests n two ways: automatcally and manually. The automatc way conssts n the use of the weghts of the mportance descrbed n the prevous secton: the documents devoted to the same mportant topcs are consdered close, even f they dffer n the detals. Alternatvely, by manually restrctng the comparson to some subtree of the concept herarchy, the user can compare the documents wth respect to a gven topc. Let us consder two documents: one mentonng the use of dogs for sabotage acts and another one mentonng the feedng of dogs. For an average user, they are moderately smlar, snce both menton dogs among other thngs. However, from the pont of vew of a zoologst, they are very smlar snce both concern wth the dogs (and not cows or crocodles). For a mltary man, though, they are not smlar at all: one of them mentons sabotage acts, and another one does not menton anythng mportant. In the latter two cases, the users can choose for the comparson the sub-trees of the topcs bology and mltary and war, correspondngly. 6 Practcal Issues The most dffcult problem n applcaton of our method s complaton of the dctonary. In our case, we compled the basc part of the Englsh herarchcal dctonary from varous sources, ncludng Roget thesaurus, WordNet, FACTOTUM SemNet and some other dctonares. However, for our goals we needed a Spansh and a French dctonary. As t turned out, automatc translaton of the Englsh dctonary gave very good results. Even a smple translaton procedure that translated the words out of context (and thus dd not dscrmnate meanngs), provded us wth a usable dctonares. Further mprovements to the Spansh and French dctonares are acheved by usng a context-senstve procedure descrbed n (Gelbuh 1999). For the languages other than Englsh, a powerful morphologcal engne has to be used to match the words n all ther morphologcal forms to the words n the documents. However, wth a smpler approach, only nouns and adectves can be used n the dctonary, that greatly reduces the complexty of morphologcal processng for the languages that do not have 288

9 Document Indexng wth a Concept Herarchy grammatcal cases. In our mplementaton, we ept both forms of Englsh, French, and Spansh nouns, as well as four forms of French and Spansh adectves, drectly n the dctonary. Fg. 6. Smlarty of the documents to a gven document wth respect to a gven aspect (topc Insttutons) 7 Dscusson and Future Wor Generally, the results obtaned n our experments show good accordance wth the classfcaton made by human experts. However, we encountered some problems wth usng our method. Most of them are related wth ambguty. Sometmes, a frequent eyword (taen out of context) proves to be mportant for a specfc topc: the noun well s an mportant term n petroleum extracton, the noun do s a term n harstyles, the noun n n poltcs, etc. However, the expresson (1) assgns too lttle weght to such eywords. To solve ths problem, we plan to add a part of speech tagger to our system. For a more detaled analyss, we wll have to add a syntactc parser to the program; however, ths would greatly slow down the system. Obvously, ths does not solve all the problems of ambguty. As we have dscussed, for the words le bll a sophstcated and not always relable algorthm s used; we plan to resolve the ambguty of ths type wth more ntellgent methods descrbed, for example, n (Gelbuh 1997). Another mportant ssue that can mprove the qualty of classfcaton s anaphora resoluton. Wth anaphorc lns at least partally resolved, the pronouns can be counted as occurrences of the correspondng nouns. 8 Conclusons We dscussed a method of document ndexng drven by a herarchcal concept dctonary. The method s statstcal-based and nvolves the quanttatve measure of the strength of the lns n the herarchy, as well as the weghts of mportance of the nodes of the herarchy for the user. We have suggested that the latter weghts depend on the database to whch the ndexng algorthm s appled. We have dscussed the automatc procedure of assgnng the correspondng weghts to the lns and the nodes n the concept herarchy. The dscussed methods have been mplemented n a system Classfer for document retreval and nvestgaton of document collectons. 289

10 Alexander Gelbuh, Grgor Sdorov and Adolfo Guzmán Arenas Though there are some problems wth the accuracy of the algorthm, the results of experments show good accordance wth the opnon of human experts. The method s practcal n the sense of nsensblty to naccuraces n the dctonary and n the sense of usng a dctonary wth a smple structure. The drectons of further mprovements to the method are related wth applcaton of ntellgent lngustc methods of lexcal and anaphorc dsambguaton. Acnowledgments The wor was partally supported by Mexcan Government (SNI, CONACyT, CGPI-IPN). References 1. Apté Ch; F. Damerau, and Sh. M. Wess, Automated learnng of decson rules for text categorzaton. ACM Transactons on Informaton Systems. Vol. 12, No. 3 (July 1994), pp Bharat K. and M. Henznger, Improved algorthms for topc dstllaton n hyper-lned envronments, 21 st Internatonal ACM SIGIR Conference, Cassdy P., An Investgaton of the Semantc Relatons n the Roget's Thesaurus: Prelmnary results, In: Proc. CICLng-2000, Internatonal Conference on Intellgent Text Processng and Computatonal Lngustcs, IPN, Mexco, 2000, Charabart S.; B. Dom, R. Agrawal, and P. Raghavan Usng taxonomy, dscrmnants, and sgnatures for navgatng n text databases, 23 rd VLDB Conference, Athenas, Greece, Cohen W. and Y. Snger, Context-senstve Learnng Methods for Text Categorzaton, Proc. of SIGIR'96, Feldman R. and I. Dagan, Knowledge Dscovery n Textual Databases, Knowledge Dscovery and Data Mnng, Montreal, Canada, Gelbuh A., "Usng a semantc networ for lexcal and syntactc dsambguaton", Proc. of Smposum Internaconal de Computacón: Nuevas Aplcacones e Innovacones Tecnológcas en Computacón, November 1997, Mexco. 8. Gelbuh A., "Syntactc dsambguaton wth weghted extended subcategorzaton frames". Proc. PACLING-99, Pacfc Assocaton for Computatonal Lngustcs, Canada, pp Gelbuh A., G. Sdorov, and A. Guzmán-Arenas, "Document comparson wth a weghted topc herarchy", Proc. 1 st Internatonal Worshop on Document Analyss and Understandng for Document Databases (DAUDD 99), 10 th Internatonal Conference and Worshop on Database and Expert Systems Applcatons (DEXA), Florence, Italy, September 1, IEEE Computer Socety Press, pp Gelbuh A., G. Sdorov, and A. Guzmán-Arenas, "A Method of Descrbng Document Contents through Topc Selecton". Proc. of SPIRE 99, Internatonal Symposum on Strng Processng and Informaton Retreval, Cancun, Mexco, September IEEE Computer Socety Press, 1999, pp Guzmán-Arenas A., Fndng the man themes n a Spansh document, Expert Systems wth Applcatons, Vol. 14, No. 1/2, Jan/Feb 1998, pp Guzmán-Arenas A., Hallando los temas prncpales en un artículo en español, Solucones Avanzadas. 1997, Vol. 5,, No. 45, p. 58, No. 49, p Hyötynem H., Text Document Classfcaton wth Self-Organzng Maps, n STeP'96, Genes, Nets and Symbols, Alander, J.; Honela, T.; Jaobsson, M. (eds.), Fnnsh Artfcal Intellgence Socety, 1996, pp Koller D. and M. Saham, Herarchcally classfyng documents usng very few words, Internatonal Conference on Machne Learnng, 1997, pp Krowetz B. Homonymy and Polysemy n Informaton Retreval, 35th Annual Meetng of the Assocaton for Computatonal Lngustcs, 1997, pp Le D.X., G. Thoma and H. Weschler, Document Classfcaton usng Connectonst Models, IEEE Internatonal Conference on Neural Networs, Orlando, FL, June 28 July 2, 1994, Vol. 5, pp Lght J., A dstrbuted, graphcal, topc-orented document search system CIKM '97, Proceedngs of the sxth nternatonal conference on Informaton and nowledge management, 1997, pp

11 Document Indexng wth a Concept Herarchy 18. Nwa Y., Sh. Nshoa, M. Iwayama, A. Taano, and Y. Ntta, Topc Graph Generaton for Query Navgaton: Use of Frequency Classes for Topc Extracton, NLPRS'97, Natural Language Processng Pacfc Rm Symposum '97, Phuet, Thaland, Dec. 1997, pp Ponte J. M. and W. B. Croft, Text Segmentaton by Topc, Frst European Conference on Research and Advanced Technology for Dgtal Lbrares, 1997, pp Seymore K. and R. Rosenfeld, Usng story topcs for language model adaptaton, Proc. of Eurospeech 97, WORDNET, Colng-ACL'98 Worshop: Usage of WordNet n Natural Language Processng Systems. August 16, 1998, Unversté de Montréal, Montréal, Canada. 291

12 Alexander Gelbuh, Grgor Sdorov and Adolfo Guzmán Arenas Alexander Gelbuh. He was born n Moscow, Russa, n He obtaned hs Master degree n Mathematcs n 1990 from the department of Mechancs and Mathematcs of the Lomonosov Moscow State Unversty, Russa, and hs Ph.D. degree n Computer Scence n 1995 from the All-Russan Insttute of the Scentfc and Techncal Informaton (VINITI), Russa. Snce 1997, he s the head of the Natural Language and Text Processng Laboratory of the Computng Research Center, Natonal Polytechnc Insttute, Mexco Cty. He s an academcan of Mexcan Academy of Scences snce 2000 and Natonal Researcher of Mexco (SNI) snce 1998; author of about 300 publcatons on computatonal lngustcs; see Grgor Sdorov. He was born n Moscow, Russa, n He obtaned hs Master degree n Structural and Appled Lngustcs n 1988 from the Phlologcal faculty of the Lomonosov Moscow State Unversty, Russa, and hs Ph.D. degree n Structural, Appled and Mathematcal Lngustcs n 1996 from the same faculty. Snce 1998, he wors for the Natural Language and Text Processng Laboratory of the Computng Research Center, Natonal Polytechnc Insttute, Mexco Cty. He s a Natonal Researcher of Mexco (SNI) snce 1999; author of about 100 publcatons on computatonal lngustcs; see Adolfo Guzmán Arenas. He was born n Ixtaltepec, Oaxaca, Méxco, n He obtaned hs Master degree n Electrcal Engneerng (Computer Scence) n 1967 and hs Ph.D. degree n Computer Scence n 1968, both from MIT, USA, under supervson of Marvn L. Mnsy. Snce 1996, he wors for the Computng Research Center, Natonal Polytechnc Insttute, Mexco, of whch he was the Drector n He s laureate of varous natonal an nsttutonal awards, ncludng the Natonal Award n Scences and Arts of Mexco, He s an ACM Fellow snce 2002, academcan of the Mexcan Academy of Scences, Mexcan Academy of Engneerng, and New-Yor Academy of Scences; Natonal Researcher of Mexco (SNI); author of numerous publcatons on computer scence; see alum.mt.edu/www/aguzman. 292