Textual Document Indexing and Retrieval via Knowledge Sources and Data Mining

Transcription

1 Textual Doument Indexing and Retrieval via Knowledge Soures and Data Mining Wesley. W. Chu, Zhenyu Liu and Wenlei Mao Computer Siene Department, University of California, Los Angeles {ww, viliu, Abstrat We present a knowledge-based query expansion tehnique to improve doument retrieval effetiveness. The general onept terms in a query are substituted by a set of speifi onept terms used in the orpus that o-our with the key query onept. Sine the expanded query mathes with the doument index terms muh better, experimental results reveal that suh query expansion produes better retrieval effetiveness than the unexpanded ones. We have also developed a new phrase-based indexing tehnique that ombines onepts with word stems. Using word stems an ompensate for the inompleteness of the knowledge soures. Experimental results reveal that using phrase-based for indexing produes more aurate doument retrieval than using word stems or onepts alone. We also present an implementation that integrates the query expansion and phrase-based indexing in a medial digital library for the retrieval of patient reords, laboratory reports and medial literature. 1. Introdution Effiient doument retrieval based on user query is ahieved by indexing. Current tehnique uses word stem to index a doument [1]. Suh tehnique suffers from the inability to math words in query with their related words suh as synonyms, hypernyms and hyponyms [2] in the douments. Therefore, there are reent attempts to index the doument based on oneptual terms. However the knowledge soures are usually inomplete. As a result, past researh reveals that although using oneptual terms for doument indexing an solve some of the problems, it annot outperform the word-stem-based model [3,4,5,6]. To remedy the inompleteness of the knowledge soures, we propose a phrase-based indexing model where we parse a doument into phrases based on the oneptual terms in domain speifi knowledge soures, and alulate the similarity between two douments using both the similarity between the onepts and the ommon word stems in them. Inluding word stem similarity in doument similarity evaluation ompensates for the inompleteness of the knowledge soures. When seeking speifi information regarding a partiular topi, the user often has to pose a general query with onept terms. For example, not knowing that ontat lenses is a treatment option for keratoonus, the user has to request for treatment options for keratoonus in the query. This results in low retrieval preision sine douments are indexed by the speifi terms. To remedy suh shortoming, we propose to substitute the general onept terms in the query with the speifi terms. The level of relevany of a speifi term in the resulting query is determined by its o-ourrene with the general onept term, whih an be mined from the orpus. Based on the query, the knowledge soures an identify the irrelevant oneptual terms and prevent them from being inserted into the query. Sine the expanded queries math better with relevant douments, the doument retrieval performane is improved. We shall first present the phrase based index tehnique and the experimental results to show the performane improvements of phrased based index over word stem and onept based index methods, Next we shall present the knowledge based query extension tehnique and present the performane improvement derived from the query extension. Finally, we present an implementation of integrating the two proposed tehniques in a medial digital library for retrieving medial textual reords and reports. 2. Phrase-based indexing To failitate disussion, we shall use the following sample query in this setion: 22 year old with hyperthermia, leukoytosis, inreased intraranial pressure, and entral herniation. Cerebral edema seondary to infetion, diagnosis and treatment. The first part of the query is a brief desription of the patient; and the seond part is the information need. 2.1 Word stem Indexing

2 A doument is ommonly represented as a vetor of terms in a vetor spae model (VSM) [1]. The basis of the vetor spae orresponds to distint terms in a doument olletion. Components of the doument vetor are the weights of the orresponding terms that represent their relative importane in the doument. In a naïvest approah, we ould treat a word as a term. Yet, morphologial variants like edema and edemas are so losely related that they are usually onflated into a single word stem, e.g., edem, by stemming [1,7]. Our sample query thus onsists of word stems hypertherm, leukoytos, inreas, intraran, pressur, et. Word stems are usually treated as notational, rather than oneptual entities. Two word stems are onsidered unrelated if they are different. For example, the stem of hyperthermia and that of fever are usually onsidered unrelated despite their apparent relationship. In stem-based VSM, word stems onstitute the basis of the vetor spae. The base vetors are orthogonal to eah other beause different word stems are onsidered unrelated. The weight w s α, u of a word stem u in a doument α is determined by the number of times u appears in α (known as the term frequeny) and the number of douments that ontain u (known as the doument frequeny) following the TF-IDF (term frequeny, inverse doument frequeny) sheme [1]. In essene, the more often u appears in α, the more important u is in α. On the other hand, the more douments u belongs to, the less disambiguating power it has, and thus the less important it is. Word stems are widely used as index terms. To improve retrieval auray, it is natural to replae word stems with onepts [3,4,5,6,8] or multiple-word ombinations [9,10]. However, previous researh showed not only no improvements, but degradation in retrieval auray when onepts were used in doument retrieval [3,4,5,6] exept when douments were very short [8]. When properly used, multiple-word ombinations were shown to improve retrieval effetiveness for some speial queries [9,10]. However, the retrieval effetiveness improvement for ad ho queries is still questionable. 2.2 Conept-based VSM Using word stems to represent douments results in the inappropriate fragmentation of onepts suh as inreased intraranial pressure into its omponent stems inreas, intraran, and pressur. Clearly, using onepts instead of single words or word stems as the vetor spae basis should produe a VSM that better mimis the human thought proesses, and therefore should result in more aurate retrieval. However, using onepts is more omplex than using word stems. First, onepts are usually represented by multi-word phrases suh as inreased intraranial pressure. More importantly, there exist synonymous and polysemous phrases. Two phrases sharing a onept are synonymous, and phrases that ould represent more than one onept are polysemous [2]. For example, hyperthermia and fever are synonymous beause they share the same onept an abnormal elevation of the body temperature. At the same time, hyperthermia is polysemous, beause in addition to the above meaning it also means a treatment in whih body tissues is exposed to high temperature to damage and kill aner ells. Synonyms an be identified with the help of a ditionary or a thesaurus. Determining whih onept a partiular polysemous phrase represents is known as word sense disambiguation (WSD) [11]. Third, some onepts are related to one another. Hypernym and hyponym relations are important oneptual relations. If we say an x is a (kind of) y then onept x is a hyponym of onept y, and y is a hypernym of x [2]. Hyperthermia is a hyponym of high body temperature; and high body temperature is a hypernym of hyperthermia. Conept identifiers are usually used to identify onepts. Using UMLS [13] as a knowledge soure, our sample query beomes (15967, ), (23518), and (151740) et., representing hyperthermia, leukoytosis, and inreased intraranial pressure, et., respetively. In onept-based VSM, the basis of the vetors spae onsists of distint onepts. To model the relationship of suh onepts as hyperthermia and elevated body temperature we remove the orthogonality onstraint on base vetors. Base vetors for two related onepts form an aute angle. Only when we annot find any reasonable relations between two onepts that we treat their orresponding vetors as orthogonal. The osine of the angle between two onept vetors is defined as the oneptual similarity between the orresponding onepts. The oneptual similarity thus ranges from 0 to 1 with 0 indiating unrelated and 1 indiating highly related onepts. To study the effets of oneptual similarities, we shall ompare two ases. In one ase, we assume all different onepts are unrelated. Therefore, all base vetors of the vetor spaes are orthogonal to one another. We label this ase as O for orthogonal. In the other ase, we derive oneptual similarities from knowledge soures. The resulting base vetors are no longer mutually orthogonal. We label this ase as NO for non-orthogonal. We derive the weight w α,xi of the i th onept x i in a doument α using a slightly modified version of TF-IDF sheme. Higher weights are assigned to longer phrases that orrespond to more speifi onepts. For example, if the term frequenies and doument frequenies for

3 inreased intraranial pressure and hyperthermia were idential, the former onept would obtain a higher weight than the latter. 2.2 Phrase-based VSM Coneptual similarities needed in onept-based VSM are derived from knowledge soures. The quality of suh VSM therefore depends heavily on the quality of the knowledge soures. The missing of ertain oneptual relations in the knowledge soures ould potentially degrades retrieval auray. For example, treating erebral edema and erebral lesion as unrelated is potentially harmful. Notiing their ommon of omponent word erebral in the above phrases, we propose phrase-based VSM to remedy the inompleteness of the knowledge soures. In phrase-based VSM, a doument is represented as a set of phrases. Eah phrase may orrespond to multiple onepts (due to polysemy) and onsist of several word stems. Our sample query now beomes [(15967, ), ( hypertherm )], [(23518), ( leukoytos )] and [(151740), ( inreas, intraran, pressur )] et. Following the TF-IDF shemes in stem-based and onept-based VSMs, we an derive the stem weight s w α, u i, k of the k th stem u i,k and the onept weight w α, x i, m of the m th onept x i,m in phrase i of α. Similar to onept-based VSM, we study two ases, O and NO. In ase O, different onepts are unrelated; while in ase NO, onepts may be related. In both ases, distint word stems are assumed to be unrelated. 2.3 Doument Similarity The similarity of two douments α and β is the osine of the angle between their orresponding doument vetors α and β ; that is, α β sim ( α, β ) = os( α, β ) = (1) α α β β We shall extend the vetor dot produtα β and denote the extended dot produt (EDP) as α β to represent the ases when the omponents of the vetors α and β are related. Using the EDP in plae of the dot produt, we derive doument similarity as, α β sim ( α, β ) = (2) α α β β EDP Derivation To derive the EDP in the phrase-based VSM, we first onsider onepts without polysemy. Thus, α β = S i, i, (3) where S i, is the oneptual ontribution of phrase i in α and phrase in β to the EDP. Assuming that eah phrase represents a single onept, we have, S ( x y ) i, = wα, x w y s i β, i, (4) where s(x i, y ) is the oneptual similarity between the i th onept x i in α and the th onept y in β. In the orthogonal ase, s(x,y) is redued to the Kroneker delta funtion, δ ( x, y) 1 = 0 x = y x y In the non-orthogonal ase, oneptual similarities are derived from knowledge soures. Coneptual Similarity, S(x,y) Given a hypernym hierarhy, the oneptual similarity s(x, y) between two onepts x and y depends on both the distane between them in the hierarhy and their generality. When two onepts are farther apart in the hypernym hierarhy, they are less similar a onept is less similarity to its grandparent than to its parent in the hypernym hierarhy. Thus we define the oneptual similarity to be inversely proportional to the number of hops between x and y, d(x,y). The generality of a onept x an be derived from the number of all its desendants D(x). The more desendants x has, the more general it is. A general onept like disease has muh more desendants than a more speifi onept like hyperthermia has. Beause of the exponential growth of the number of desendants when a onept moves up a tree struture, we take the logarithm of the number of desendant in oneptual similarity alulation. The oneptual similarity is therefore defined to be inversely proportional to the logarithm of the number of desendants of the two. A final onsideration is the boundary ase when we reah the leaves of the hypernym tree. Let us assume we have two onepts x o and y o, where x o is the only diret hypernym of y o, y o is the only hyponym of x o, and y o has no hyponym of its own. Conepts x o and y o are so muh alike that we define the oneptual similarity between them to be lose to 1, say 0.9, to represent suh loseness. As a result, the oneptual similarity between onepts x and y is, s ( x, y) = d ( x, y) log ( 1+ D( x) + D( y) ) 2 (5) In order to use (4) in the presene of polysemy, we need to disambiguate senses. To avoid WSD ost, we use the most popular onept that a phrase represents as the meaning of the phrase. Alternatively, we derive the

4 oneptual ontribution to the similarity between two phrases using an aggregation of (4) over all possible onept pairs, where eah pair onsists of one onept from eah phrase. The ontribution of word stems to the EDP is the sum of the weight produt for those word stems ommon to both phrases, s s s Si, = wδ, u w v ( ui k v l ) i k, l,,, θ δ,, (6) k, l where u i,k and v,l are the k th word stem in phrase i in α and l th word stem in phrase in β respetively. Given the ontribution of onepts and stems, (4) and (6), we selet the larger of the two as the ontribution of phrase i in α and phrase in β to the EDP and get, s α β = max( S i,, Si, ) (7) i, Suh seletion remedies the inompleteness of the knowledge soures. α α and β β an be derived similar to (7). The doument similarity an then be omputed from (2) using these EDPs Experimental Results The Test Colletion, OHSUMED OHSUMED [12] is a large test olletion used in many information retrieval system evaluations. The test set onsists of a referene olletion, a query olletion, and a set of relevane udgments. The referene olletion is a subset of the MEDLINE database. Eah referene ontains a title, an optional abstrat, a set of MeSH headings, author information, publiation type, soure, a MEDLINE identifier, and a sequene identifier. The query olletion onsists of 106 queries. Eah query ontains a patient desription, an information request, and a sequene identifier. The sample query we use in this paper is query 57 in the olletion. 14,430 referenes out of the 348K are udged by human experts to be not relevant, possibly relevant, or definitely relevant to eah query.we use the title, the abstrat, and the MeSH headings to represent eah doument; and the patient desription, and the information request to represent eah query. The Knowledge Soure UMLS [13] is a medial lexial knowledge soure and a set of assoiated lexial programs. The knowledge soure onsists of UMLS Metathesaurus, SPECIALIST lexion, and UMLS semanti network. Espeially of interest to us is its entral voabulary omponent the Metathesaurus. It ontains biomedial phrases from more than 60 voabularies and lassifiations. The Metathesaurus ontains 1.6M phrases representing over 800K onepts. A onept unique identifier (CUI) identifies eah onept. UMLS tends to assign a smaller CUI to a more popular sense of a phrase. Therefore, we use the onept with the smallest CUI in oneptual ontribution alulation (2). Our experiment results show that suh heuristi produes retrieval auray omparable to that produed by the aggregation approah where we onsider all oneptual similarities due to different sense ombinations from the phrases. The Metathesaurus enodes many oneptual relations. We onentrate on hypernym relations. Two relations in UMLS roughly orrespond to the hypernym relations: the RB (border than) and the PAR (parent) relations. For example, hyperthermia has a parent onept body temperature hange. We ombine the 838K RB and 607K PAR relations into a single hypernym hierarhy. Hypernymy is transitive [14]. For example, sign and symptom is a hypernym of body temperature hange and body temperature hange is a hypernym of hyperthermia, so sign and symptom is also a hypernym of hyperthermia. However UMLS Metathesaurus enodes only the diret hypernym relations but not the transitive losure. We derive the transitive losure of the hypernym relation and use (5) to alulate the oneptual similarities. Phrase Detetion Given a set of douments (106 queries and 14K udged douments of OSHUMED), we need to detet any ourrenes in a set of phrases (1.3M phrases in UMLS). We adopt the Aho-Corasik algorithm [15] for the set-mathing problem to detet phrases: First, Aho-Corasik algorithm detets all ourrenes of any phrase in a doument. But we only keep the longest, most speifi phrase. For example, although both edema and erebral edema are deteted in the sample query, we keep only the latter and ignore the former. Seond, to detet multi-word phrases, we math stems instead of words in a doument with UMLS phrases. We use Lovins stemmer [7] to derive word stems. To avoid onflating different abbreviations into a single stem, we define the stem for a word shorter than four haraters to be the original word. Third, stop-word removal is performed after the multi-word phrase detetion. In this way, we orretly detet seondary to and arinoma from erebral edema seondary to arinoma. We would inorretly detet seondary arinoma if the stop-words ( to in this ase) were removed before the phrase detetion. Disussion of Results To alulate retrieval auray using preision-reall [1], we ombined the possibly relevant and definitely relevant udgments in OHSUMED into a single relevant ategory. Based on the type of VSM, we alulate the doument similarity between eah of the 14K douments and eah of the 105 queries (one query does not have relevant doument). For a given VSM

5 and a query, we rank the douments from the most to the least similar to the query. When a ertain number of douments are retrieved, preision is the perentage of retrieved douments that are relevant; and reall is the perentage of the relevant douments that has been retrieved so far. We evaluate the retrieval auray by interpolating the preision values at eleven reall points. The overall effetiveness of different VSM is then ompared by averaging over the performane of all the 105 queries (Figure 1). The average of the eleven preision values gives an overview of the effetiveness of eah VSM Stems, O Conepts, O Conepts, NO Phrases, O Phrases, NO Figure 1. 1 Comparison of the average preisionreall over 105 queries. 1. The baseline labeled (Stems, O) uses stem-based VSM. Its 11-point average preision is Considering the ontribution of onepts only, and treating different onepts as unrelated (Conepts, O), we arrive at an 11-point average preision of 0.260, whih is a 28% derease from the baseline. 3. Similar to 2, but taking the onept inter-relationship into onsideration (Conepts, NO), we ahieve a signifiant improvement over 2. The average auray is similar to that of the baseline. 4. Considering ontribution of both onepts and word stems in a phrase, but treating different onepts as unrelated (Phrases, O), we arrive at an 11-point average preision of 0.375, a 3% improvement over the baseline. 5. Similar to 4, but taking onept interrelations into onsideration (Phrases, NO), we ahieve an 11-point average preision of 0.420, whih is a signifiant 16% improvement over the baseline. Our experiment results reveal that viewing douments as onepts only and treating different onepts as unrelated an ause the retrieval auray to deteriorate (ase 2). Considering onept inter-relations (ase 3) or relating different phrases by their shared word stems (ase 4) an both improve retrieval auray. The extended dot produt ombines ontributions from the onepts and word stems. The phrase-based VSM utilizes suh extended dot produt and yields signifiant improvement in retrieval auray. 3. Enhane retrieval performane via query expansion When posing a query, a user usually has a main obetive (key onept, key ) in mind and uses additional general supporting oneptual terms, s, to speify ertain aspets of key. For example, when a user asks Keratoonus, treatment options., Keratoonus (an eye disease) is the key onept whereas treatment options is a general supporting onept. Although suh query is easy to form, it does not math well with relevant douments that use suh speifi supporting onepts as ontat lens or keratoplasty. To remedy this problem, we propose to substitute the general supporting onepts by speifi onepts that used in the relevant douments. We need to selet the set of speifi onepts, and determine the weight of eah of these onepts. We shall first present the weight determination method, and then ompare two onept seletion approahes. For a speifi onept, its weight should represent the degree of orrelation between and the key onept term, key. For example, ontat lens is a treatment option for Keratoonus but not Bak pain, and therefore it should assign a larger weight in the expansion of Keratoonus, treatment options than that of Bak pain, treatment options. We shall now present a salable method for suh weight assignment. We shall first represent onepts into inverted doument vetors, and then use the similarity between the two inverted doument vetors to represent the orrelation between the two onepts. Given a orpus of n douments, the inverted doument vetor for onept i, θ, is defined as an n i dimensional vetor. The weight of omponent i in the vetor represents the term frequeny of onept i in eah doument. For example, if a orpus ontains douments D 1, D 2 and D 3, and onept ours three, zero, and two times in these douments respetively, then θ =<3, 0, 2>. i We further define the orrelation between onepts i and as: θ θ i orrel( i, ) = os ( θ, θ ) i = θ θ i θ i θ The orrelation between two onepts ranged from 0 to 1. For example, if θ i =<3, 0, 2>, θ =<6, 0, 4>, and θ k =<0, 1, 0>, then orrel( i, )=1 and orrel( i,

6 k )=0. The orrelation between all pairs of onepts an be omputed offline and stored in a onept orrelation table (see Figure 3). For query expansion, the weight assigned to i is the orrelation between a supporting onept i and key. There are two onept seletion approahes: with or without knowledge soures. When no knowledge soure is available, all the supporting onept terms that have zero orrelation with the key onept, key, an be viewed as irrelevant onepts and filtered out. Let 1, 2,..., k be all the onepts that have nonzero orrelation with key. The expanded query beomes <( key,1), ( 1, orrel( 1, key )),, ( k, orrel( k, key ))> In this expanded query vetor, the higher the orrelation value orrel( i, key ) is, the more emphasis is plaed on the vetor omponent of i. Clearly, the weight assigned to key is 1 sine orrel( key, key ) = 1. Sine a general supporting onept in a query usually only relevant to one or two aspets of key, this motivate the use of knowledge soure together with the key onept and the general supporting onept in the query to selet the relevant set of speifi supporting onepts. For example, "Keratoonus, treatment options" is emphasizing on the treatment options for the disease, instead of diagnosis methods or auses for the disease. ULMS [13] indiates only three ategories of medial onepts as potential treatments: "Therapeuti and Preventive Proedures", "Medial Devies" and "Pharmalogial Substane". Thus only onepts that belong to these three ategories will be expanded into the query. By filtering out the irrelevant supporting onepts, the omputation omplexity is greatly redued. In addition, the preision in the low reall region inreases. To evaluate the performane improvement of query expansion, we selet 28 OHSUMED queries [12] that ontain general supporting onepts as the test set. Our experimental results for the query set reveal the average query expansion size without knowledge soure is 1227 terms per query, while using ULMS, the average expansion size redued to 82.3 terms per query. This represents more than an order of magnitude redution in query expansion size. The retrieval performane improvement for the set of expaned OHSUMED queries is shown in Figure 2. We note that the expansion queries performed better than the non-expansion ases (base line). The query derived from knowledge-based expansion performs better than the ases without knowledge base in the low reall region, and the performane is reverse in the high reall region. This is beause the non-knowledge based expansion ase inludes many irreverent speifi onepts, whih resulted in low preision in the low reall region. Sine relevant douments in the high reall region are not typially overed by the highly orrelated supporting onepts in the query, adding more terms by the non-knowledge based method yields preision better preision in that region. Aside from the omputation omplexity saving of using the knowledge based expansion method, the hoie of whih expansion method to use depends on the appliation. When low reall region is more of onern, then knowledge-based method is preferable; and if high reall region is more important, then the non-knowledge based method should be onsidered reall Figure 2. Retrieval performane improvements with query expansion 4. Appliations Single Word Stem Baseline Non-knowledge Based Expansion Knowledge Based Expansion We shall now present a doument retrieval system that integrates the knowledge-based query expansion and phrase-based indexing for a digital medial library. As shown in Figure 3, the system onsists of three subsystems: a doument indexing engine, a query expansion engine, and a doument retrieval engine. The doument indexing engine proesses the knowledge soure and the orpus offline and generates data struture neessary for the online query expansion and doument retrieval. When the system reeives a user query, the query expansion engine expands the query and the doument retrieval engine then returns a set of douments relevant to the user. The medial knowledge soure, UMLS, is used in the system for phrase detetion, oneptual similarity derivation, and expansion term filtering. The doument indexing engine proesses UMLS and the douments orpus separately. The oneptual similarity alulator derives the oneptual similarities between onepts from UMLS and stores them in the oneptual similarity table. The phrase detetor identifies onepts in UMLS from the douments. The inverse doument frequeny alulator uses the output of the phrase detetor to onstrut an inverse doument frequeny table. The phrase weight alulator in turn alulates the weights of onepts and word stems in eah phrase, and onverts the original orpus into a phrase-indexed orpus. The orrelation alulator then derives the oneptual

7 Original Query Phrase Detetor for Query Domain Speifi Knowledge Soure Query Expansion Engine General Term Detetor List of General Supporting Conepts Has General Supporting Conepts? N Y Query Expander C 1 C 2 C 1 C 2 Conept Correlation Table Expanded Query Correlation Calulator Corpus Phrase-indexed Corpus Phrase Weight Calulator for Douments Phrase Detetor for Douments Phrase Weight Calulator for Query Inverse Doument Frequeny Table Inverse Doument Frequeny Calulator Doument Similarity Calulator and Ranker Doument Retrieval Engine C 1 C 2 C 1 C 2 Coneptual Similarity Table Coneptual Similarity Calulator Doument Indexing Engine Douments Returned Figure 3. A phrase based indexing and query expansion doument retrieval system. orrelations from the phrase-indexed orpus and stores them in the onept orrelation table. When the system reeives a user query, phrase detetor first parse the query into phrases. The general term detetor then determines if any query expansion is neessary by onsulting the general supporting onept list. If no general supporting onepts are deteted. The original query is diretly input into the doument retrieval engine. Otherwise, the query expander replaes general onept terms with the speifi ones in the appropriate ategories as speified in UMLS. The weights of the expanded onepts are looked up from the onept orrelation table. The doument retrieval engine ompares the expanded query with the phrase-index and returns a set of douments to the user. First, the similarity alulator onsults the oneptual similarity table to alulate the phrase-based doument similarities. The ranker then returns to the user those douments the most similar to the query. 5. Conlusion We introdued a knowledge-based tehnique to rewrite a user query ontaining general oneptual terms to one ontaining speifi terms that related to the general supporting onept. These speifi supporting terms are mined from the orpus. Experimental results show that retrieval using suh expanded queries is more effetive than the original queries. Beause the additional onept terms inluded in the expanded query are seleted from few ategories as indiated by the knowledge soure, the average size of the expanded queries in our approah is muh smaller (less than an order of magnitude) than that produed by the full query

8 expansion tehnique and also yield better retrieval performane in the low reall region whih is of interest to most appliations. We developed a new vetor spae model that uses phrases to represent douments. Eah phrase onsists of multiple onepts and words. Similarity between two phrases is ointly determined by the oneptual similarity and their ommon word stems. Our experimental result reveals that the phrase-based VSM yields a 16% inrease of retrieval effetiveness than the stem-based VSM. This improvement is beause multiword onepts are natural unit of information and using word stems in phrase-based doument similarity ompensates for the inauray in oneptual similarities derived from inomplete knowledge soures. We also presented an implementation that integrates the above tehniques into a digital medial library at UCLA for the retrieval of patient reords, laboratory reports and medial literatures. [13] National Library of Mediine. UMLS Knowledge Soures, 12 th edition, 2001 [14] J. Lyons. Semantis, 1977 [15] A.V. Aho and M.J. Corasik. Effiient String Mathing: an Aid to Bibliographi Searh. In CACM, 18(6), , 1975 Referenes [1] G. Salton and M.J. MGill. Introdution to Modern Information Retrieval, 1983 [2] G.A. Miller, R. Bekwith, C. Fellbaum, D. Gross and K. Miller. Introdution to WordNet: an On-line Lexial Database. In WordNet: an Eletroni Lexial Database, 1-19, 1998 [3] M. Mitra, C. Bukley, A. Singhal and C. Cardie. An Analysis of Statistial and Syntati Phrases. In Pro. RIAO97, , 1997 [4] R. Rihardson and A.F. Smeaton. Using WordNet in a Knowledge-based Approah to Information Retrieval. In Pro. 17 th BCS-IRSG, 1995 [5] M. Sussna. Text Retrieval using Inferene in Semanti Matanetworks. PhD Thesis, University of California, San Diego, 1997 [6] E.M. Voorhees. Using WordNet to Disambiguate Word Sense for Text Retrieval. In Pro. 16 th ACM-SIGIR., , 1993 [7] J.B. Lovins. Development of a Stemming Algorithm. In Mehanial Translation and Computational Linguistis, 11(1-2), 11-31, 1968 [8] A.F. Smeaton and I. Quigley. Experiments on using Semanti Distanes Between Words in Image Caption Retrieval. In 19 th Pro. ACM-SIGIR, , 1996 [9] D. Johnson, W.W. Chu, J.D. Dionisio, R.K. Taira and H. Kangarloo. Creating and Indexing Teahing Files from Free-text Patient Reports. In AMIA 99, 1999 [10] J.A. Goldman, W.W. Chu, D.S. Parker and R.M. Goldman. Term Domain Distribution Analysis: A Data Mining Tool for Text Databases. In 2001 IMIA Yearbook of Medial Informatis, , 2001 [11] N. Ide and J. Véronis. Word Sense Disambiguation: the State of the Art. In Computational Linguistis, 24(1), 1-40, 1998 [12] W. Hersh, C. Bukley, T.J. Leone and D. Hikam. OHSUMED: an Interative Retrieval Evaluation and New Large Test Colletion for Researh. In Pro. 22 nd ACM- SIGIR Conf., , 1994