Automatically attaching semantic metadata to web services

Automatically attaching semantic metadata to Web services Andreas HeßNicholas KushmerickComputer Science Department,University College Dublin,Ireland{andreas.hess,nick}@ucd.ieAbstractEmerging Web standards promise a network of het-erogeneous yet interoperable Web Services.WebServices would greatly simplify the developmentof many kinds of data integration and knowledgemanagement applications.Unfortunately,this vi-sion requires that services provide large amountsof semantic metadata“glue”.As afirst step to au-tomatically generating such metadata,we describehow machine learning and clustering techniquescan be used to attach attach semantic metadata toWeb forms and services.1IntroductionEmerging Web standards such as UDDI[],SOAP [/TR/soap],WSDL[/TR/wsdl]and DAML-S [/services]promise an ocean of Web Services, networked components that can be invoked remotely using standard XML-based protocols.For example,significant e-commerce players such as Amazon and Google export Web Services giving public access to their content databases. The key to automatically invoking and composing Web Services is to associate machine-understandable semantic metadata with each service.While the details are beyond the scope of this paper,the various Web standards involve metadata at various levels of abstraction,from high-level ad-vertisements that facilitate discovering relevant services,to low-level input/output specifications of particular operations.A central challenge to the Web Services initiative is the lack of tools to(semi-)automatically generate the necessary metadata.In this paper we explore the use of machine learn-ing techniques to automatically create such metadata from training data.Such an approach complements existing uses of machine learning to facilitate the Semantic Web,such as for information extraction[6;8;2]and for mapping between heterogeneous data schemata[4].Specifically,we describe and evaluate supervised learning techniques for attaching semantic metadata to Web forms and fields(Sec.2),as well as unsupervised clustering techniques for discovering semantically related Web Services(Sec.3).2Web form classificationProblem formulation.Web form instances are structured objects:a form comprises one or morefields,and eachfield in turn comprises one or more terms.More precisely,a form F i is a sequence offields,written F i=[f1i,f2i,...],and each field f j i is a bag of terms,written f j i=[t j i(1),t j i(2),...]. We assume two taxonomies for attaching semantic meta-data to forms andfields.First,we assume a domain taxonomy D.Domains capture the overall purpose of a form,such as“searching for a book”,“finding a job”,“querying a airline timetable”,etc.We use S MALL-C APS to indicate domains,so we might have D= {S EARCH B OOK,F IND J OB,Q UERY F LIGHT,...}. Second,we assume a datatype taxonomy T.Datatypes do not relate to low-level encoding issues such as“string”or “integer”,but rather to the expected semantic category of a field’s data,such as“book title”,“salary”,“destination air-port”,etc.SansSerif style indicates datatypes,so we might have T={BookTitle,Salary,DestAirport,...}.The Web form learning problem is as follows.The input is set of labelled forms andfields;that is,a set{F1,F2,...}of forms together with a domain D i∈D for each form F i,and a datatype T j i∈T for eachfield f j i∈F i.The output is a form classifier;that is,a function that maps an unlabelled form F i,to a predicted domain D i∈D,and a predicted datatype T ji∈T for eachfield f ji∈F i.Generative model.Our solution to the Web form classi-fication is based on a stochastic generative model of a hy-pothetical“Web service designer”creating a Web page to host a particular service.First,the designerfirst selects a domain D i∈D according to some probability distribution Pr[D i].For example,in our Web form data described in rel-ative to forms forfinding colleges,so Pr[S EARCH B OOK] Pr[F IND C OLLEGE].Second,the designer selects datatypes T j i∈T appropriate to D i,by selecting according to some distribution Pr[T j i|D i]. For example,presumably Pr[BookTitle|S EARCH B OOK] Pr[DestAirport|S EARCH B OOK],because services forfinding books usually involve a book’s title,but rarely involve air-ports.On the other hand,Pr[BookTitle|Q UERY F LIGHT] Pr[DestAirport|Q UERY F LIGHT].Finally,the designer writes the Web page that implements the form by coding eachfield in turn.More precisely,for each selected datatype T j i,the designer uses terms t j i(k)drawn ac-cording to some distribution Pr[t j i(k)|T j i].For example,pre-sumably Pr[title|BookTitle] Pr[city|BookTitle],be-cause the term title is much more likely than city to occur in afield requesting a book title.On the other hand,pre-sumably Pr[title|DestAirport] Pr[city|DestAirport]. Parameter estimation The learning task is to estimate the parameters of the stochastic generative model from a set of training data.The training data comprises a set of N Web forms F={F1,...,F N},where for each form F i the learn-ing algorithm is given the domain D i∈D and the datatypesT j i of thefields f j i∈F i.The parameters to be estimated are the domain probabili-tiesˆPr[D]for D∈D,the conditional datatype probabilities ˆPr[T|D]for D∈D and T∈T,and the conditional termprobabilitiesˆPr[t|T]for term t and T∈T.We estimate these parameters based on their frequency in the training data:ˆPr[D]=NF (D)/N,ˆPr[T|D]=M F(T,D)/M F(D),andˆPr[t|T]=WF (t,T)/W F(T),where N F(D)is the numberof forms in the training set F with domain D;M F(D)is the total number offields in all forms of domain D;M F(T,D)is the number offields of datatype T in all forms of domain D; W F(T)is the total number of terms of allfields of datatype T;and W F(t,T)is the number of occurrences of term t in all fields of datatype t.Classification.Our approach to Web form classification in-volves converting a form into a Bayesian network.The net-work is a tree that reflects the generative model:there is a root node representing the form’s domain,children represent-ing the datatype of eachfield,and grandchildren encoding the terms used to code eachfield.In more detail,a Web form to be classified is converted into a three-layer tree-structured Bayesian network as fol-lows.Thefirst(root)layer contains just a single node do-main that takes on values from the set of domains D.The second layer consists of one child datatype i of domain for eachfield in the form being classified,where each datatype i take on values from the datatype set T.The third(leaf)layer comprises a set of children {term1i,...,term K i}for each datatype i node,where K is the number of terms in thefield.The term nodes take on val-ues from the vocabulary set V,defined as the set of all terms that have occurred in the training data.Fig.1illustrates the network that would be constructed for a form with threefields and K terms for eachfield.(Each field contains the same number K of terms/field for simplic-ity;in fact,the number of term nodes reflects the actual num-ber of terms in the parentfield.)The conditional probability tables associated with each node correspond directly to the learned parameters men-tioned earlier.That is,Pr[domain=D]≡ˆPr(D), Pr[datatype i=T|domain=D]≡ˆPr(T|D),and Figure1:The three-layer tree-structured Bayesian network used to classify a form containing threefields.Pr[term k i=t|datatype i=T]≡ˆPr(t|T).Note that the conditional probabilities tables are identical for all datatype nodes,and for all term nodes.Given such a Bayesian network,classifying a form F i= [f1i,f2i,...]involves“observing”the terms in eachfield(i.e., setting the probability Pr[term k i=t j i(k)]≡1for each term t ji(k)∈f ji),and then computing the maximum-likelihood form domain andfield datatypes consistent with that evi-dence.Evaluation We have evaluated our approach using a col-lection of129Web forms comprising656fields in total,for an average of5.1fields/form.As shown in Fig.2,the domain taxonomy D used in our experiments contains6domains,and the datatype taxonomy T comprises71datatypes.The forms were manually gathered by manually brows-ing Web forms indices such as for relevant forms.Each form was then inspected by hand to assign a domain to the form as a whole,and a datatype to eachfield. After the forms were gathered,they were segmented into fields.We discuss the details below.For now,it suf-fices to say that we use HTML tags such as<input>and <textarea>to identify thefields that will appear to the user when the page is rendered.After a form has been seg-mented intofields,certain irrelevantfields(e.g.,submit/reset buttons)are discarded.The remainingfields are then assigned a datatype.Afinal subtlety is that somefields are not easily interpreted as“data”,but rather indicate minor modifications to either the way the query is interpreted,or the output presentation. For example,there is a“help”option on one search services that augments the requested data with suggestions for query refinement.We discarded suchfields on a case-by-case basis;a total of12.1%of thefields were discarded in this way. Thefinal data-preparation step is to convert the HTML fragments into the“form=sequence offields;field=bag of terms”representation.The HTML isfirst parsed into a sequence of tokens.Some of these tokens are HTMLfield tags(eg.,<input>,<select>,<textarea>).The form is segmented intofields by associating the remaining tokensDomain taxonomy D and number of forms for each domainS EARCH B OOK(44)F IND C OLLEGE(2)S EARCH C OLLEGE B OOK(17)Q UERY F LIGHT(34)F IND J OB(23)F IND S TOCK Q UOTE(9)Datatype taxonomy T(illustrative sample)Address NAdults Airline Author BookCode BookCondition BookDetailsBookEdition BookFormat BookSearchType BookSubject BookTitle NChildren CityClass College CollegeSubject CompanyName Country Currency DateDepartDateReturn DestAirport DestCity Duration Email EmployeeLevel...Figure2:Subsets of the domain and datatype taxonomies used in the experiments.with the nearestfield.For example,“<form>a<inputname=f1>b c<textarea name=f2>d</form>”would be segmented as“a<input name=f1>b”and“c<textarea name=f2>d”.The intent is that this segmentation process will associatewith eachfield a bag of terms that provides evidence of thefield’s datatype.For example,our classification algorithm will learn to distinguish labels like“Book title”that are as-sociated with BookTitlefields,from labels like“Title(Dr, Ms,...)”that indicate PersonTitle.Finally,we convert HTML fragments like“Enter name: <input name=name1type=text size=20><br>”that correspond to a particularfield,into thefield’s bag of terms representation.We process each fragment as follows. First,we discard HTML tags,retaining the values of a set of“interesting”attributes,such as an<input>tag’s name attribute.The result is“Enter name:name1”.Next, we tokenize the string at punctuation and space characters, convert all characters to lower case,apply Porter’s stemming algorithm,discard stop words,and insert a special symbol encoding thefield’s HTML type(text,select,radio-button, etc).This yields the token sequence[enter,name,name1, TypeText].Finally,we apply a set of term normalizations, such as replacing terms comprising just a single digit(letter) with a special symbol SingleDigit(SingleLetter),and delet-ing leading/trailing numbers.In this example thefinal result is the sequence[enter,name,name,TypeText]. Results We begin by comparing our approach to two simple bag of terms baselines using a leave-one-out methodology. For domain classification,the baseline uses a single bag of all terms in the entire form.For datatype classification,the baseline approach is the naive Bayes algorithm over its bag of terms.For domain prediction,our algorithm has an F1score of 0.87while the baseline scores0.82.For datatype predic-tion,our algorithm has an F1score of0.43while the baseline scores0.38.We conclude that our“holistic”approach to form andfield prediction is more accurate than a greedy baseline approach of making each prediction independently.While our approach is far from perfect,we observe that form classification is extremely challenging,due both to noise in the underlying HTML,and the fact that our domain and datatype taxonomies contain many classes compared to tradi-tional(usually binary!)text classification tasks.While fully-automated form classification is our ultimate goal,an imperfect form classifier can still be useful in inter-active,partially-automated scenarios in which a human gives0.40.50.60.70.80.910.20.30.40.50.60.70.80.91F1degree of assistanceα33333Figure3:Domain prediction F1as a function of the fraction αoffields’datatypes supplied by the user.the domain or(some of)the datatypes of a form to be labelled, and the classifier labels the remaining elements.Ourfirst experiment measures the improvement in datatype prediction if the Bayesian network is also provided as evi-dence the form’s domain.In this case our algorithm has an F1score of0.51,compared to0.43mentioned earlier.Our second experiment measures the improvement in do-main prediction if evidence is provided for a randomly chosen fractionαof thefields’datatypes,for0≤α≤1.α=0cor-responds to the fully automated situation in which no datatype evidence is provided;α=1requires that a person provide the datatype of everyfield.As shown in Fig.3,the domain classification F1score increases rapidly asαapproaches1. Our third investigation of semi-automated prediction in-volves the idea of ranking the predictions rather than requir-ing that the algorithm make just one prediction.In many semi-automated scenarios,the fact that the second-or third-ranked prediction is correct can still be useful even if thefirst is wrong.To formalize this notion,we calculate F1based on treating the algorithm as correct if the true class is in the top R predictions as ranked by posterior probability.Fig.4shows the F1score for predicting both domains and datatypes,as a function of R.R=1corresponds to the cases described so far.We can see that relaxing R even slightly results in a dramatic increase in F1score.So far we have assumed unstructured datatype and domain taxonomies.However,domains and datatypes exhibit a nat-ural hierarchical structure(eg,“forms forfinding something”vs.“forms for buying something”;or“fields related to book information”vs.“fields related to personal details”).It seems reasonable that in partially-automated settings,predicting a similar but wrong class is more useful than a dissimilar class. To explore this issue,our research assistants converted their domain and datatype taxonomies into trees,creating ad-ditional abstract nodes to obtain reasonable and compact hier-archies.We used distance in these trees to measure the“qual-0.40.50.60.70.80.9112345678910F 1rank threshold Rform domain 3333333field datatype+++++++++++Figure 4:F1as a function of rank threshold R .ity”of a prediction,instead of a binary “right/wrong”.For domain predictions,our algorithm’s prediction is on average 0.40edges away from the correct class,while the baseline al-gorithm’s predictions are 0.55edges away.For datatype pre-diction,our algorithm’s average distance is 2.08edges while the baseline algorithm averages 2.51.As above,we conclude that our algorithm outperforms the baseline.3Web Service clusteringClustering.As a second approach towards our goal of au-tomatically creating Web Services metadata,we explored the use of unsupervised clustering algorithms to automatically group services into semantically related categories.Note that a given Web Service can export more than one operation.Each operation would correspond to a single Web form.Let C ={C 1,C 2,...}be the categories discovered by clustering a collection of Web Services.Then each C i corresponds for-mally to a subset of the domain ontology D :C i ∈2D for each i .Ultimately,we intend to use the category C i associ-ated with a particular Web Service as additional top-down ev-idence that could be exploited by the Bayesian network used to classify each of its operations.Clustering in information retrieval has been applied pri-marily on unstructured documents.As Web Service descrip-tions are structured documents,we developed a clustering al-gorithm that takes advantage of this structure.UDDI.As Web Services become more common,there is an increasing need for search and discovery tools.UDDI reg-istries promise to solve the discovery problem,but UDDI’s search capabilities are rather weak.UDDI is more like the Domain Name Service than a search engine:it provides an abstraction over the low level details by avoiding hard-coded access addresses,because applications can query the UDDI registry for technical details given a Web Service name.On the other hand,UDDI has richer features than a Web search engine,such as support for the classification of Web Services into a taxonomy like the North American Industry Classification System (NAICS).These taxonomies provide a means of classifying a Web Service as a whole.This is still far less than a complete semantic annotation,and the situation is complicated due to multiple competing taxonomies.Some-times not even a part-whole relationship is exploited when searching the registry for a service that belongs to a certainindustry classification.For instance,a Web Service that ad-vertised itself as a credit card service might not be found if the user searches for a banking service,although it might be exactly what the user wants.Creating the technical descriptions of Web Services such as WSDL is more or less an automatic procedure,creating semantic descriptions—even merely classification into an in-dustry taxonomy—involves additional work by the provider.Our goal is to provide a semi-automated tool to simplify this burden,by using clustering techniques to discover a set C of categories that correspond to,for example,an industry taxon-omy like NAICS.Web Service corpus.For our clustering experiments,we focused on UDDI data structures.However,we obtained a collection of 425Web Services not through a UDDI registry,but from ,a WSDL indexing service.As shown in Fig.5,the information available is a Web Service description in the WSDL format and the name and descrip-tion of the service.(We did not use the service provider as a text source.)We parsed the port types,operations and mes-sages from the WSDL and extract names and comments.We did not extract standard XML schema data types like string or integer.These Web Services were than manually classified into a hierarchical taxonomy.To avoid bias,the person who clas-sified the services was a research student with no previous experience with Web Services.The person had the same in-formation as given on ,and was allowed to inspect the WSDL description if necessary.The person was advised to adaptively create new classes while classifying the Web Services and was allowed to arrange the classes in a hi-erarchy.As shown in Fig.6,the 425Web Services were clas-sified into 24top level classes.The Web Services were not evenly distributed over the 24classes:some top level cate-gories contain only two Web Services while others contain over 50services in several subclasses.Clustering algorithms.We tested four clustering algo-rithms on our collection of Web Services:a hierarchical group average algorithm (e.g.[11]or [10]),the Word-IC al-gorithm [12],a variant of the group average clusterer that we call Common-Term,and a clusterer that exploits the struc-tured nature of the data we have gathered.The Common-Term algorithm differs from the standard group average clustering in the way the centroid document vector is computed.Instead of using all terms from all the sub-clusters,only the terms that occur in all sub-clusters form the centroid.Like the Word-IC algorithm,our hope is that this leads to short and concise cluster labels.For each of the three unstructured clustering algorithms,we convert a Web Service A into a bag of words text(A ).text(A )contains all the information about a particular ser-vice.We use the Porter stemming algorithm and a stop-word list to reduce the dimensionality of the task,and we use the standard cosine approach to measuring the similarity σ(a,b )between two “documents”a and b .For the first three unstruc-tured clustering algorithms,we define the similarity σ(A,B )Cartoon Communication Converter Country Information Courier Information Database Provider Data Management Developers Engineering Finder Flights GamesGenbank Graphics License Mathematics Money MusicNews Parser Poll Creation Server Info User Groups WebFigure6:Web Service categories C(only top-level categories areshown).Figure5:Data structure for our Web Service corpus. between two services A and B as the similarity of their bags of words:σ(A,B)=σ(text(A),text(B)).The structured clusterer measures similarity in terms of both the similarity of the Web Services and the descriptions, and also the similarity between the input/output messages. The similarity between messages of two Web Services A and B is defined as the maximum similarity between all messages msgs(A)of thefirst service and all messages msgs(B)of the other service.The intuition is that similar services tend to not only have similar textual descriptions,but also similar names for their messages.The combined similarityσ(A,B) is defined as the product between the textual similarity of the service and the similarity of the messages:σ(A,B)=σ(text(A),text(B))·maxn∈msgs(A)m∈msgs(B)σ(n,m),where for the structured clusterer text(A)excludes terms in the messages associated with service A’s operations.The structured clusterer is still work in progress,as we are currently experimenting with different ways to exploit the structured nature of Web Services.For example,other ways of combining the similarities,like using a weighted average, are possible.It is also possible to cluster on the messagesfirst and then use the cluster distance of the messages as a simi-larity measure.The disadvantage of the latter is that every service can have dozens of messages,and so this preliminary clustering step is infeasible without additional heuristics. Evaluation.We compared the four clustering algorithms with the manually-generated reference clusters mentioned above,using three ways to measure the similarity between two clusters over a set of objects.Thefirst measure was in-troduced by Zamir et al[12].This measure looks at a cluster in the target distribution and classifies each pair of services as a true positive or a false positive.The second similarity measure is the conditional probabil-ity of two Web Services being in the same cluster in our refer-ence classification,given they are in the same machine gener-ated cluster and vice versa,and the conditional probability of two Web Services being in different clusters in the reference given they are in different clusters in the machine clustering and vice versa.To get an overall measure,we multiplied these probabilities.The probability of two Web Services being in the same ref-erence class given they are in the same cluster alone is similar to the precision in traditional information retrieval,so we also report this value.This measure is not affected by unclustered services,while Zamir’s measure explicitly penalises unclus-tered services.Fig.7shows the quality of the clusters generated compared to the reference clusters,according to the three evaluation metrics,for each of the four algorithms.Not surprisingly, none of the algorithms does particularly well,because the clustering problem is quite challenging.In many cases even humans disagree on the correct classification.For example, manually organizes Web Services into their own taxonomy,and thefinal column in Fig.7shows that these reference clusters bear little resemblance to ours.Further-more,we have24categories in our reference classification, which is a rather high number.Our structured algorithm is penalised by Zamir’s metric, because it produces many unclustered services.The reason for this is that the similarity between two documents in the Structured algorithm reaches zero,if either the text or the messages have zero similarity.Thus,some Web Services will never be clustered in the structured algorithm.While the clusters do not agree well with our reference clusters,anecdotally wefind that the discovered clusters do generally make intuitive sense.An exception concerns syn-onyms:all algorithms make mistakes clustering Web Ser-vices containing the word“quote”,where“quote of the day”and“stock quote”services are inadvertently merged.The Common-Term and Structured clustering algorithms achieve a high precision,but create very many top level clusters,and, in the case of the Structured clusterer,a very high number ofWord-IC Common-Term Group Average Structured SALCentral Zamir’s quality measure-0.4738-0.2339-0.5390-0.3960-0.4163 Combined probabilities0.02230.02150.03430.02850.0329 Precision0.10240.33170.12570.41360.1774 Number of unclustered services0721170Number of top level clusters147285214Figure7:Cluster quality compared to reference cluster;see below for a discussion of the SALCentral column.unclustered documents.The Common-Term algorithm per-forms also best when looking at Zamir’s quality measure,al-though all algorithms only reach negative values.4Discussion.Future work.We are currently extending our classification and clustering algorithms in several directions.For exam-ple,our approaches ignore valuable sources of evidence—such as the actual data passed to/from a Web Service—and it would be interesting to incorporate such evidence into our algorithms.Our clustering algorithm could be extended in a number of ways,such as using statistical methods such as latent semantic analysis as well as thesauri like WordNet to improve the discovery and clustering of Web Services.We are currently integrating the clustering and classifica-tion algorithms.As described earlier,the clusters can provide additional evidence for classification,in that clustering on the Web Service level can give hints on the domains of the ser-vice’s operations.Although the classification approach has been evaluated only on HTML forms,we anticipate that the method can be extended to real Web Services.Related work.There has been some work on semantic matching of Web Services[9;1],but they require semantic metadata such as DAML-S.Clustering is a well-known technique in information re-trieval,although it has not yet been applied to Web Ser-vices.Besides the traditional group-average or single-link approaches,newer algorithms like the mentioned Word-IC or the Scatter/Gather-algorithms[3]exist.However,these algo-rithms have been applied mostly to unstructured data.The search capabilities of UDDI are still very restricted,different extensions are available or under development(e.g.UDDIe [/user/A.Shaikhali/uddie]).Kerschberg et al are planning to apply the techniques they introduced in Web-Sifter[5]to UDDI.When we actually want to simultaneously invoke multiple similar Web Services and aggregate the results,we encounter the problem of XML schema mapping.This problem is ad-dressed in different ways in[4;7].Conclusions.The emerging Web Services protocols rep-resent exciting new directions for the Web,but interoper-ability requires that each service be described by a large amount of semantic metadata“glue”.We have presented two approaches to automatically generating such metadata,and evaluated our approach on a collection of Web Services and forms.Although we a far from being able to fully automatically create semantic metadata,we believe that the methods we have presented here are a reasonablefirst step.Our pre-liminary results indicate that some of the requisite semantic metadata can be semi-automatically generated using machine learning,information retrieval and clustering techniques. Acknowledgments.This research was supported by grants SFI/01/F.1/C015from Science Foundation Ireland,and N00014-00-1-0021from the US Office of Naval Research. References[1]J.Cardoso.Quality of Service and Semantic Composition ofWorkflows.PhD thesis,University of Georgia,2002.[2] F.Ciravegna.Adaptive information extraction from text byrule induction and generalization.In17th International Joint Conference on Artifical Intelligence,2001.[3] D.Cutting,J.Pedersen, D.Karger,and J.Tukey.Scat-ter/gather:A cluster-based approach to browsing large doc-ument collections.In Proceedings of the Fifteenth Annual In-ternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval,pages318–329,1992.[4] A.Doan,P.Domingos,and A.Halevy.Reconciling schemasof disparate data sources:A machine-learning approach.In Proc.SIGMOD Conference,2001.[5]L.Kerschberg,W.Kim,and A.Scime.Intelligent web searchvia personalizable meta-search agents.pages1345–1358, 2002.[6]N.Kushmerick.Wrapper induction:Efficiency and expres-siveness.Artificial Intelligence,118(1–2):15–68,2000. [7]S.Melnik,H.Molina-Garcia,and E.Rahm.Similariyflood-ing:A versatile graph matching algorithm.In Proc.of the International Conference on Data Engineering(ICDE),2002.[8]I.Muslea,S.Minton,and C.Knoblock.A Hierachical Ap-proach to Wrapper Induction.In Proc.3rd Int.Conf.Au-tonomous Agents,pages190–197,1999.[9]M.Paolucci,T.Kawamura,T.Payne,and K.Sycara.Semanticmatchmaking of web services capabilities.In Int.Semantic Web Conference,2002.[10]G.Salton and M.McGill.Introduction to Modern InformationRetrieval.McGraw-Hill,1983.[11] C.J.van rmation Retrieval.Butterworths,Lon-don,2nd edition,1979.[12]Oren Zamir,Oren Etzioni,Omid Madani,and Richard M.Karp.Fast and intuitive clustering of web documents.In Knowledge Discovery and Data Mining,pages287–290,1997.。