7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ http://www.esi.uem.es/~jmgomez/tutorials/ecmlpkdd02/ 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ 287/,1( 1 Text Mining and Internet Content Filtering José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 2 Goals and methodology * An overview of Text Mining ... * ... exploring two Internet content filtering applications ... * ... following the standard KDD process and ... * ... working with operational code 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 3 Outline I 1. Text Mining: What is it and what is it not? 2. Learning from text when we know what about to learn 3. Learning from text when we do not know what about to learn 4. Tools for Text Mining Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 4 Outline II 5. Application to the detection of offensive websites 6. Application to the detection of unsolicited bulk email 7. Challenges in Text Mining 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ 7(;7 0,1,1* 1 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 1 Outline 1. Introduction 2. Definition of Text Mining 3. Applications of Text Mining 4. Problems with Textual Data 5. Text Mining and KDD: the process 6. Content Based Text Processing Tasks 7. Text Processing Basics 8. Case Study: CORA Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 2 Introduction I * Attractive field due to the Internet / Intranet / Digital Libraries explosion * Increasing amount of text in electronic form * E.g. [Moore 00], by July ­ Number of unique pages on Internet: 2.1 billion ­ Unique pages added per day: 7.3 million * E.g. [Oracle 97] ­ text represents about 90% of all information handled by an organization 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 3 Introduction II * About the name(s) ­ Text Mining, Text Data Mining, Knowledge Discovery in Text, Knowledge Discovery in Textual Data(bases) ­ Like Data Mining as a step in the KDD process [Fayyad et al. 96], we could see the Text Mining step in the Knowledge Discovery in Textual Data process ­ We will take the sense by Hearst [Hearst 99] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 4 Definition of Text Mining I * Several definitions in the literature (e.g. [Dörre et al. 99, Feldman & Dagan 95, Hearst 99, Kodratoff 99, Rajman & Besaçon 98]) * Extending the KDD definition, Text Mining is "WKH QRQWULYLDO H[WUDFWLRQ RI LPSOLFLW SUHYLRXVO\ XQNQRZQ DQG SRWHQWLDOO\ XVHIXO LQIRUPDWLRQ IURP ODUJH DPRXQWV RI WH[WXDO GDWD" 3 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 5 Definition of Text Mining II * To what extent is something SUHYLRXVO\ XQNQRZQ? ­ In Hearst's opinion, nor even the writer knows => real new knowledge = real Text Mining * E.g. The discovery of an DEVROXWHO\ QHZ, potentially effective treatment for a disease by exploring scientific literature * E.g. The discovery of the fact that private and not public funding leads to more inventions by exploring patent files ­ Others think we have to rediscover the information the author encoded in text Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 6 Definition of Text Mining III * But to get knowledge, users must be helped to locate, examine and relate suitable information ... * ... through text analysis, classification and understanding tasks * So, Text Mining is not Information Access but relies on it 4 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 7 Applications I * The same as KDD applications, but working with textual data ­ marketing manufacturing ­ financial investment health care ­ decision support fraud detection ­ science etc. * We review some examples Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 8 Applications II * Knowledge Management [Semio 02] ­ Enormous need to manage and control great quantities of textual and other information that drive businesses ­ Knowledge management is "WKH SURFHVV RI FDSWXULQJ D FRPSDQ\ V FROOHFWLYH H[SHUWLVH ZKHUHYHU LW UHVLGHVLQ GDWDEDVHV RQ SDSHU RU LQ SHRSOH V KHDGVDQG GLVWULEXWLQJ LW WR ZKHUHYHU LW FDQ KHOS SURGXFH WKH ELJJHVW SD\RII." 5 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 9 Applications III * For instance, TAKMI (Text Analysis and Knowledge MIning) by IBM [Nasukawa & Nagano 01] * Mining textual databases in PC help centres, they can ­ automatically detect product failures ­ determine issues that have led to rapid increases in the number of calls and their underlying reasons ­ analyse help centre productivity and changes in customers' behaviour involving a particular product Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 10 Applications IV * Personal Intelligent Information Access Assistants [Mladenic 99] * Help users to access (find, relate) information from several sources and to turn it into knowledge through, for instance ­ gathering and processing text ­ finding relations among pieces of information ­ providing suitable metaphors for information browsing 6 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 11 Applications V * For instance, Personal WebWatcher [Mladenic 99] ­ A content-based intelligent agent that uses textlearning for user-customized web browsing ­ Note that the text writer may not be aware of some information consumers needs ­ For instance, the agent can be used to track competitors or clients web pages Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 12 Applications VI * Predicting trends on the basis of textual evidence [Grobelnik et al. 00] * For instance, EAnalyst [Lavrenko et al. 00] ­ Discovers the trends in time series of numeric data and attempts to characterize the content of textual data that precede these events ­ The objective is to use this information to predict the trends in the numeric data based on the content of textual documents that precede the trend ­ E.g. To predict the trend in stock prices based on the content of new articles published before that trend occurs 7 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 13 Problems with Textual Data I * The known KDD problems and challenges [Fayyad et al. 96] extend to Textual Data ­ Large (textual) data collections ­ High dimensionality ­ Over fitting ­ Changing data and knowledge ­ Noisy data ­ Understand ability of mined patterns ­ etc... Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 14 Problems with Textual Data II * But there are new problems ­ Text is not designed to be used by computers ­ Complex and poorly defined structure and semantics ­ But much harder, DPELJXLW\ * In speech, morphology, syntax, semantics, pragmatics * For instance, intentionality ­ Multilingualism * Lack of reliable and general translation tools 8 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 15 Text Mining and KDD: the process I * We will focus on Text Mining techniques and subordinate text tasks * But under the KDD standard process * Text analysis, processing tasks play different roles in different steps Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 16 Text Mining and KDD: the process II * The standard KDD process (borrowed from [Fayyad et al. 96]) 9 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 17 Text Mining and KDD: the process III Categorization Information Retrieval POS Tagging Word Sense Disambiguation Partial Parsing Term Clustering Clustering Summarization Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 18 Text Mining and KDD: the process IV * Or the text-related task may be previous to a KDD process, as Information Extraction (IE) [Grobelnik et al. 00] * IE aims at filling domain dependent templates with data buried in text items * E.g. WebKB aims at probabilistic, symbolic knowledge base that mirrors the content of the World Wide Web [Ghani et al. 00] 10 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 19 Content Based Text Processing Tasks I * Taxonomy of Text Mining subtasks based on [Lewis 92] * Dimensions ­ Size of text ­ Involve supervised or unsupervised learning ­ Text classification vs. understanding * Assigning documents or parts to a number of groups vs. * More complex access to document content * Note it is not a sharp division Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 20 Content Based Text Processing Tasks II * Sample text classification tasks Document Clustering, Topic Detection and Tracking Latent Semantic Indexing, Automatic Thesaurus Construction, Key Phrase Extraction Unsupervised learning Text Categorization, Filtering, Topic Detection and Tracking POS Tagging, Word Sense Disambiguation Supervised learning DocumentsWords 11 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 21 Content Based Text Processing Tasks III * Sample text understanding tasks SummarizationWord Sense DiscoveryUnsupervised learning Information ExtractionSupervised learning DocumentsWords Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 22 Text Processing Basics I * Most text processing tasks involve a number of steps, including ­ Text instances representation (indexing in Information Retrieval) ­ Learning ­ Evaluation and presentation * Those steps can be seen as KDD steps over text data ­ For instance, text representation corresponds to selection, processing and transformation steps 12 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 23 Text Processing Basics II * Text instances representation ­ The goal is to capture text semantics ­ A requirement is to avoid intensive manual processing (hand-coding) except for text labelling ­ Involves * Feature definition * Feature selection and extraction ­ This step reduces concept space dimensionality Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 24 Text Processing Basics III * Feature definition ­ Text is usually represented as a "bag of words" ­ Which in fact is a the Information Retrieval (IR) Vector Space Model [Salton 89] ­ A text sequence is represented as a concept weight vector ­ Concepts are words, word stems, word phrases ­ Weights can be binary, frequency-based 13 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 25 Text Processing Basics IV * Feature definition ­ Semantic similarity between natural language expressions is captured through the cosine formula q d2 d1 d3 = = = = m 1i m 1i 2 ki 2 ji m 1i kiji kj wqwd wqwd )q,d(sim Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 26 Text Processing Basics V * Feature definition ­ A very frequent representation involves ­ Filtering according to a high frequency (stop) word list to discard low-semantics words (prepositions, etc) [Salton 89] * BNC frequency lists, IR stop-lists ­ Stemming words to achieve a canonical concept representation (e.g. analysis, analysing, analyser are collapsed to ANALY) * Porter stemmer for English 14 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 27 Text Processing Basics VI * Feature definition ­ Concept weights are often WILGI kind [Salton 89] ­ tf(i,j) is the number of times that concept i occurs in document j of the text collection ­ N is the number of documents in the text collection ­ df(i) is the number of documents in which concept i occurs = )i(df N log)j,i(tf)j,i(W 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 28 Text Processing Basics VII * Feature selection ­ A subset of the original concepts is extracted to avoid low representative concepts [Sebastiani 02] ­ If supervised learning follows, information theoretic or statistical measures are used to rank concepts according to their quality ­ Measures can be global (when quality for all classes is measured) or local (concepts are specific to each class) 15 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 29 Text Processing Basics VIII * Feature selection ­ Some effective quality metrics include * Information Gain - IG (locally defined) * Being ti the ith concept and ck the kth class in the text collection * Require text items labelled with class identifiers = }c,c{x }t,t{y 2 kk ii )y(P)x(P )y,x(P log)y,x(P)k,i(IG Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 30 Text Processing Basics IX * Feature selection * Document Frequency (DF) is the number of documents in which the concept occurs * Does not require text items labelled with class identifiers * But DF can also be defined according to classes ( )ki ctP)k,i(DF = 16 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 31 Text Processing Basics X * Feature selection * Several more including odds ratio, 2 [Sebastiani 02] * Variable effectiveness * For instance, for Text Categorization [Yang & Pedersen 97] ­ IG and 2 are very effective (allow to eliminate 99% of concepts without effectiveness decrease in classification) ­ DF is quite effective (90% elimination) ­ Mutual Information and Term Strength are bad Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 32 Text Processing Basics XI * Feature extraction ­ Based on the assumption that original concepts are not very representative (because of ambiguity, lack of statistical evidence) ­ A new concept set is produced by assorted procedures including * Thesaurus construction * Latent Semantic Indexing * Concept Indexing * Phrase construction 17 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 33 Text Processing Basics XII * Feature extraction ­ Thesaurus construction [Salton 89] * Using co-occurrence statistics to detect semantic regularities across concepts * E.g. Class #764 in a engineering text collection is UHIXVDO UHIXVDO GHFOLQLQJ QRQFRPSOLDQFH UHMHFWLRQ GHQLDO ­ Also known as Term Clustering [Lewis 92] ­ Related to Latent Semantic Indexing and Concept Indexing Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 34 Text Processing Basics XIII * Feature extraction ­ Latent Semantic Indexing [Deerwester et al. 90, Dumais 95] ­ A set of document vectors indexed according a set of concepts is transformed to reduce the number of concept dimensions ­ A mapping function is obtained by applying a singular value decomposition to the matrix formed by the original document vectors ­ Address synonymy and polysemy 18 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 35 Text Processing Basics XIV * Feature extraction ­ Concept Indexing ­ By using a semantic net or ontology of concepts ­ For instance, the Lexical Database WordNet in [Gonzalo et al. 98] ­ Faces a problem of ambiguity ­ An automatic, effective Word Sense Disambiguation process is required ­ But may be language independent using EuroWordNet [Vossen 98] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 36 Text Processing Basics XV * Feature extraction ­ Phrase construction * Concepts sequences can be recorded (n-grams) and added to the concept set * Alternatively, some Natural Language Processing can be applied to build linguistically motivated phrases (noun phrases) ­ For instance, using Part-Of-Speech Tagging, shallow parsing and regular expressions * Mixed results [Lewis 92, Riloff & Lehnert 94] 19 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 37 Text Processing Basics XVI * Learning ­ Supervised learning algorithms * Most are general Machine Learning (ML) algorithms including decision tree learners, rule learners, neural networks and Support Vector Machines (SVM), (Naive) Bayesian approaches, linear function learners, instance-based classification, etc * Some specific algorithms like Rocchio from IR * Maybe the most effective are SVMs ­ Unsupervised learning algorithms * Again most coming from ML including hierarchical clustering methods, Expectation-Maximization, k-means, etc Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 38 Text Processing Basics XVII * Direct evaluation ­ Supervised classification * Standard metrics coming from IR and ML including Recall, Precision, Accuracy, Error, F, etc * Statistical tests not often used ­ Unsupervised classification * Comparing with a manual classification * Entropy, etc 20 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 39 Text Processing Basics XVIII * Direct evaluation in Supervised classification tnfnC fptpC CCSystem &RQWLQJHQF\ PDWUL[ Actual fntp tp )r(recall + = fptp tp (p)precision + = tnfpfntp tntp accuracy +++ + = r 1 )1( p 1 1 F -+ = rp pr2 F1 + = Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 40 Text Processing Basics XIX * Indirect evaluation ­ The task A is a part of a more complex task B ­ Approaches for task A are compares as they affect task B effectiveness ­ For instance, a Word Sense Disambiguation approach can be evaluated as it affects an automatic translation approach 21 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 41 Text Processing Basics XX * Visualization ­ From standard visualization tools from KDD... * Including decision tree graph tools, etc ­ To specific metaphors designed for text presentation (in Human-Computer Interaction, Information Access, etc) * Including those by Hearst and others ­ In Hearst opinion, visualization is the core of Text Mining systems [Hearst 99], because TM is user/application centric Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 42 Text Processing Basics XXI * Decision Tree in Mine Set by Silicon Graphics 22 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 43 Text Processing Basics XXII * Clusters presentation in ThemeScapes [Wise et al.95] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 44 Text Processing Basics XXIII * Hierarchical presentation in Cat-A-Cone [Hearst & Karadi 97] 23 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 45 Text Processing Basics XXIV * Hypertext graphics in Mapuccino (formerly WebCutter) [Maarek & Shaul 97] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 46 Text Processing Basics XXV * Query word frequency in TileBars [Hearst 95] 24 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 47 Case Study: CORA I * CORA is a publicly available search engine on computer science research papers [McCallum et al. 00] * Available at http://www.cora.whizbang.com/ * Built using a number of text processing techniques * It can be used for computer scientific knowledge discovery Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 48 Case Study: CORA II 25 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 49 Case Study: CORA III * Integrates a number of techniques and tasks ­ Intelligent spidering the web for computer science research papers using reinforcement learning ­ Text Categorization to automatically classify documents into a topic hierarchy, using Naive Bayes and Expectation-Maximization ­ Information Extraction for the identification of titles, authors, etc using Hidden Markov Models ­ Information Retrieval for accessing documents and citation analysis for ranking according impact Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 50 Summarizing * We will review a number of content based text processing tasks * The goals are ­ to describe them and introduce their techniques ­ to show their role in Text Mining ­ to analyse one of them (text categorization) as a KDD process itself 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ /($51,1* :+(1 .12:,1* 1 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 1 Outline 1. Introduction 2. Text Categorization 3. Text Filtering 4. Topic Detection and Tracking 5. Part Of Speech Tagging 6. Word Sense Disambiguation 7. Shallow Parsing 8. Information Extraction Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 2 Introduction * Supervised learning tasks ­ Textual data manually labelled with class values ­ Possibility of using concept quality metrics in text representation ­ Automatic induction of automatic classifiers by Machine Learning (ML) algorithms ­ Hit-mistake evaluation metrics (precision, recall, etc) 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 3 Text Categorization I * Text categorization (TC) is the automatic assignment of documents to a set of predefined classes * Classes are usually content based (topics, keywords, subject headings) but can also be genres, authors, etc * Documents are e-mails, reports, books, web pages, news items, etc Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 4 Text Categorization II * The dominant approach is ­ Given a set of manually classified (labelled) documents ­ Use IR and ML techniques to induce an automatic classifier or new documents * This way, the knowledge acquisition of knowledge based classifiers is alleviated * See [Sebastiani 02] for an in-depth survey 3 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 5 Text Categorization III * Applications ­ Knowledge management (automatic document organization for knowledge sharing) ­ Document indexing in libraries ­ Web page classification into Internet directories ­ We will later focus on harmful Internet document identification for filtering (VSDP e-mail, pornographic web pages) ­ Many others including author identification, automatic essay grading, language guessing, etc Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 6 Text Categorization IV * The basic model involves ­ Documents content representation as bags of words with stop words filtering and word stemming ­ Feature selection and extraction ­ Learning a classifier using some ML algorithm including the full range of options ­ Evaluating the effectiveness of the learned classifier 4 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 7 Text Categorization V * Document representation ­ Riloff describes experiments in which stemming negatively affects performance [Riloff 95] ­ Document structure is rarely used, except for hypertext * For instance, in [Attardi et al. 99] HTML documents are represented by their blurb (hyper linked text pointing to the document) * Also, web pages can be classified using only hyperlink net structure, detecting hubs and authorities [Chakrabarti et al. 98] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 8 Text Categorization VI * Document representation ­ The use of phrases (either statistical or linguistic) as concepts for representation has reported assorted results * Failure reported in [Lewis 92] * Moderate success reported in [Tzeras & Hartmann 93] * Currently pursued [Caropreso et al. 01, Mladenic & Grobelnik 98] 5 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 9 Text Categorization VII * Feature selection ­ The most basic approach is deleting low frequency terms ­ A wide range of statistical quality metrics * Concepts are selected according their predictive value * Include IG, 2, DF, etc discussed above ­ The most effective are IG, 2 and DF according to [Yang & Pedersen 97] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 10 Text Categorization VIII * Feature selection ­ Interestingly, class dependent metrics can be averaged over all classes ­ Given a metric denoted by X(t,c), being t a concept and c a class in a set C, several possible averages including = Cc avg )c,t(X)c(P)t(X )}c,t(X{max)t(X Cc max = 6 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 11 Text Categorization IX * Feature extraction ­ Latent Semantic Indexing can be considered a positive technique (see [Sebastiani 02]) ­ Concept indexing based on taxonomies like the lexical database WordNet is * Successful for IR (e.g. [Gonzalo et al. 98]) * Unsuccessful for TC [Junker & Abecker 97, Scott & Matwin 99] * But WordNet can be successfully used in TC [Buenaga et al. 00, Urea et al. 01, Benkhalifa et al. 01] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 12 Text Categorization X * Machine learning classifiers ­ Nearly all methods and algorithms have been applied to the task ­ Most effective approaches include * Support Vector Machines (e.g. [Dumais et al. 98, Drucker et al. 99]) * K-Nearest Neighbours (e.g. [Yang 99, Larkey 99]) * AdaBoost-ed C4.5 (e.g. [Schapire & Singer 00]) 7 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 13 Text Categorization XI * Machine learning classifiers ­ Support Vector Machines ­ The goal is finding a surface that separates the positives from the negatives by the widest possible margin ­ The SVM method chooses the middle element from the "widest" set of parallel hyper planes in the N-dimensional space (being N the number of indexing concepts) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 14 Text Categorization XII * Machine learning classifiers ­ Support Vector Machines * Positive (+) and negative (o) instances * 2-dimensional space * Detecting the most important instances for separating rest of examples * Called support machines * Instances need not to be linearly- separable * Separating surfaces need not to be hyper planes Borrowed from [Sebastiani 02] 8 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 15 Text Categorization XIII * Machine learning classifiers ­ Support Vector Machines ­ Good effectiveness ­ Fast training for linear SVM (which result in linear classification functions) ­ Feature reduction not required * Robust to over fitting Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 16 Text Categorization XIV * Machine learning classifiers ­ k-Nearest Neighbours ­ A king of example/instance based classification, or memory based learning ­ Classifying a new instance using the classes of known instances ­ Voting classes of k neighbours according to distance to the new instance ­ Several distances available (cosine, dot product, Euclidean) 9 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 17 Text Categorization XV * Machine learning classifiers ­ k-Nearest Neighbours Comparison with centroid-based classification, borrowed from [Sebastiani 02] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 18 Text Categorization XVI * Machine learning classifiers ­ k-Nearest Neighbours ­ Very effective in classification ­ Not very efficient (but indexing techniques are now web scale, see Google) 10 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 19 Text Categorization XVII * Machine learning classifiers ­ Boosting ­ Combining a set (committee) of same-kind classifiers successively learned by a weaker method (e.g. C4.5) ­ Next classifier is induced mainly on instances misclassified by previous classifiers ­ Classification is based on weighted vote of all learned classifiers ­ Good results in BOOSTEXTER (AdaBoost+C4.5) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 20 Text Categorization XVIII * Evaluation ­ Mainly concerned with effectiveness, less with efficiency ­ Standard IR & ML metrics presented above (recall, precision, accuracy, F1, etc) ­ In multi class situations, at least report F1 by * Macro averaging ­ averaging on the number of classes * Micro averaging ­ computing over all decisions at once 11 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 21 Text Categorization XIX * Evaluation ­ Cross-validation is not frequent ­ Some available test collections include * Reuters-21578 * The Reuters Corpus Volume 1 * OHSUMED * 20-NewsGroups * Ling-Spam Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 22 Text Categorization XX * Evaluation ­ Scarce statistical testing (intro in [Yang & Liu 99]) ­ Accuracy and error do not fit well TC because class distribution is usually highly biased ­ Now an increasing use of cost-sensitive metrics for specific tasks (e.g. weighted accuracy, ROCCH method [Gomez 02]) 12 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 23 Text Categorization XXI * Interesting work in TC ­ Using unlabelled data in TC (e.g. [Nigam et al. 00]) ­ Using Yahoo-like hierarchical structure (e.g. [Mladenic 98]) ­ Using other text features (e.g. [Forsyth 99, Kessler et al. 97, Sahami et al. 98, Gómez et al. 00]) * Linguistic-like in genre or author identification, VSDP classification Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 24 Text Filtering I * Text Filtering (TF) is an information seeking process in which documents are selected from a dynamic text stream to satisfy a relatively stable and specific information need * E.g. News items from newspapers are daily collected and delivered to a user according his/her interests = the personalized newspaper * See e.g. [Oard & Marchionini 96] for overview 13 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 25 Text Filtering II * Also known as Selective Delivery of Information (SDI) * Systems use IR, ML and User Modelling techniques to induce and refine a user model which is used to select new documents from the stream * Collaborative vs. content based filtering Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 26 Text Filtering III * Product recommendation in Amazon (content) ­ According a personal profile accounting for * A set of categories (DVD, Computer games, Music) and subcategories (genres) ­ Starting with preferred items * Authors, titles, brands ­ Recommendation of new releases ­ Of course it is not text-content based, but on the purchasing history 14 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 27 Text Filtering IV * Product recommendation in Amazon (collaborative or social) ­ According to other customers purchases "&XVWRPHUV ZKR ERXJKW WKLV ERRN DOVR ERXJKW" ­ Based on * previous annotations by other users * and generating a user segmentation * A trend is combining both ideas (see e.g. [Good et al. 99]) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 28 Text Filtering V * Applications ­ CRM & marketing (e.g. cross-selling, recommendation) ­ Information delivery at organizations for Knowledge Management ­ Information access assistants (Personal WebWatcher [Mladenic 99]) ­ Filtering news items in Newsgroups ­ In Text Mining, allows to personalize information access to the current knowledge discovery task 15 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 29 Text Filtering VI * The process in content-based filtering ­ The filter usually starts at least with * A set of selected documents or * A set of user-defined keywords ­ A basic user model is induced from that information ­ The user model is refined according to relevance judgements from the user Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 30 Text Filtering VII * The basic model involves ­ Documents content representation as bags of words with stop words filtering and word stemming ­ Feature selection ­ Learning and updating a classifier usually from relevance feedback in IR ­ Evaluating the effectiveness of the learned classifier 16 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 31 Text Filtering VIII * Following [Allan 96] ­ The user initially describes his/her information interest as a standard keyword-based query ­ Documents come in batches, and after each arrival the user provides relevance judgements ­ The user model is refined using judged documents, but * It is assumed the system cannot store all documents permanently * Original user concepts must be retained to avoid erroneous bias Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 32 Text Filtering IX * Following [Allan 96] ­ After each batch is judged, 100 top scoring concepts are added for the next cycle * Concepts in documents from the new batch are first ordered by UWI (# times they occur in positive documents) * Top 500 concepts are re-ranked according to the Rocchio formula * Where weights (w) are computed in a WILGI fashion * And top 100 terms are added (always retaining user's) relnonrelqueryi w 2 1 w2wRocchio --+= 17 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 33 Text Filtering X * Following [Allan 96] ­ A precision-based evaluation shows that adding concepts in batches is nearly as effective as adding all concepts at once ­ Note that concepts are never removed from the profile ­ Other representations and algorithms are possible * E.g. [Bloerdon et al. 96] use thesaurus categories as concepts * E.g. [Tan & Teo 98] use a neural-network-kind algorithm Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 34 Topic Detection and Tracking I * Topic Detection and Tracking (TDT) is the automatic detection of novel events in chronologically-ordered streams of news stories, and tracking these events over time * Events the user want to detect are not know previously by him/her, and so retrieval is inadequate * TDT is an application by itself * See [Yang et al. 99] for an overview 18 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 35 Topic Detection and Tracking II * Events (USAir-417 crash) are not topics (airplane accidents) and must be separated * TDT is probably the purest Text Mining task since we want to discover new events * TDT approaches combine several kinds of learning including ­ supervised (TC) for tracking ­ and unsupervised (Clustering) for detection Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 36 POS-Tagging I * Part Of Speech Tagging (POS-Tagging) is labelling each word in a sentence with its appropriate part of speech ­ E.g. The-AT representative-NN put-VBD chairs-NNS on-IN the-AT table-NN ­ Where AT = Determiner-article, NN = Noun-singular, VBD = Verb-past tense, NNS = Noun-plural, IN = Preposition * Words show limited POS ambiguity * See [Manning & Schütze 99] for an overview 19 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 37 POS-Tagging II * There are several tag sets (Brown tag set, Penn Treebank tag set, etc) ranging in granularity, complexity, etc ­ E.g. 56 tags in the Penn Treebank ­ E.g. 197 tags in the London-Lung Corpus * Granularity directly affects performance Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 38 POS-Tagging III * POS-Tagging makes sense as an intermediate task for others * E.g. with shallow parsing ­ For creating linguistically motivated indexing terms (retrieval, categorization, filtering, etc) ­ For detecting slot fillers candidates in Information Extraction ­ For detecting answer candidates in Question Answering 20 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 39 POS-Tagging IV * A range of approaches ­ Markov Models (e.g. [Church 88, DeRose 88]) ­ Hidden Markov Models (e.g. [Jelinek 85, Cutting et al. 92a]) => the most popular ­ Transformation-based learning [Brill 95] ­ Decision trees [Mrquez et al. 00, Schmid 94] ­ Etc Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 40 Word Sense Disambiguation I * Most words in natural language have several meanings or senses * E.G. Bank in WordNet 1. depository financial institution, bank, banking concern, banking company -- (a financial institution that accepts deposits and channels the money into lending activities; "he cashed a check at the bank"; "that bank holds the mortgage on my home") 2. bank -- (sloping land (especially the slope beside a body of water); "they pulled the canoe up on the bank"; "he sat on the bank of the river and watched the currents") ... 10. bank -- (a flight maneuver; aircraft tips laterally about its longitudinal axis (especially in turning)) 21 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 41 Word Sense Disambiguation II * Word Sense Disambiguation (WSD) is to determine which of the senses of an ambiguous word is invoked in a particular use of the word * Usually, WSD is stated as ­ Having a set of word senses candidates listed or defined in a dictionary, thesaurus, etc ­ Detecting the most suitable sense of a word among them given the context of usage * Good overview in [Manning & Shütze 99], Chapter 7 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 42 Word Sense Disambiguation III * Again, WSD makes sense as an intermediate task for other text processing tasks, including ­ Machine translation (because each sense may result in a different translation) ­ IR and TC (because the sense of a word is highly influential in the relevance or adequacy as predictor of the word) ­ Spelling correction (for instance, to to determine when diacritics should be inserted) 22 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 43 Word Sense Disambiguation IV * Two kind of methods ­ Dictionary based (running from [Lesk 86] and [Yarowsky 92] to [Agirre & Rigau 96]) ­ Training corpus based (from [Gale 92] and [Mooney 96] to [Pedersen 02]) * A trend is the combination of techniques (e.g [Stevenson & Wilks 99]) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 44 Word Sense Disambiguation V * Dictionary based WSD ­ Information sources are * The text context of occurrence of the word to disambiguate * The available information for the senses in the dictionary ­ Hence, it is no supervised ­ The basic approach is comparing (e.g. computing the overlap) between both information sources 23 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 45 Word Sense Disambiguation VI * Dictionary based WSD ­ A very interesting approach is that in [Agirre & Rigau 96] ­ Information about senses is collected from WordNet considering semantic relations and population of the hierarchy ­ The system tries to resolve the lexical ambiguity of nouns by finding the combination of senses from a set of contiguous nouns that maximises the Conceptual Density among senses Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 46 Word Sense Disambiguation VII * Training based WSD ­ Information sources are * The text context of occurrence of the word to disambiguate * The senses in the dictionary * A manually disambiguated corpus (e.g. Semcor) ­ A supervised learning based classifier is trained on the corpus to enable predictions on new words 24 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 47 Word Sense Disambiguation VIII * Training based WSD ­ An interesting study is that in [Mooney 96] ­ The task is disambiguating the word "line" manually tagged according to its six Wordnet senses ­ For instances representation, the words occurring in the precedent and current sentences, filtered according a stoplist and further stemmed are taken as binary features Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 48 Word Sense Disambiguation IX * Training based WSD [Mooney 96] ­ A number of learning approaches are compared including a Bayesian classifier, a perceptron, a decision tree learner, kNN, two rule learners and a decision list learner ­ The training data is highly biased to the "product" sense (it occurs more that five times than others) ­ Bayes and the perceptron perform best according to precision 25 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 49 Shallow Parsing I * Also named Robust Parsing, Chunk Parsing and Chunking [Abney 91, Vergne 00, Tjong & Buchholz 00] * Given a sentence, the goal is to find a partial parsing of it, in which non-overlapping phrases and the relations among them are identified * E.g. "He reckons the current account deficit will narrow to only # 1.8 billion in September." [NP He ] [VP reckons ] [NP the current account deficit ] [VP will narrow ] [PP to ] [NP only # 1.8 billion ] [PP in ] [NP September ] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 50 Shallow Parsing II * Very related to POS Tagging ­ POS Tagging is nearly a requirement for chunking ­ Sometimes, the same methods are used for both tasks (e.g. Hidden Markov Models) ­ Both tasks are syntactic annotation at very close levels of complexity (neither of them capture natural language recursive nature) ­ In fact, they are combined for the applications, that includxe sophisticate indexing for text classification and finding slot fillers in Information Extraction 26 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 51 Shallow Parsing III * It is worth studying the chunking evaluation at the Computational Natural Language Learning Workshop 2000 * The Penn Treebank was processed to convert full parses into chunk sequences * There were 211727 tokens and 106978 chunks, where 55081 (51%) were noun phrases, 21467 (20%) were verb phrases and 21281 (20%) were prepositional phrases Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 52 Shallow Parsing IV * A number of approaches were tested ­ Rule based systems hand coded from scratch or adapted from a full parser (3) ­ Memory based systems (1) ­ Statistical methods including Markov Models, Hidden Markov Models and maximum entropy methods (4) ­ Combined approaches with committees built over a variety of base learners (SVMs, memory based, etc) 27 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 53 Shallow Parsing V $SSURDFK 3UHFLVLRQ 5HFDOO ) Combined SVM 93.45% 93.51% 93.48 Combined WPDV & MBL 93.13% 93.51% 93.32 Combined MBL 94.04% 91.00% 92.50 Hidden Markov Models 91.99% 92.25% 92.12 Rules 91.87% 92.31% 92.09 Maximum Entropy 92.08% 91.86% 91.97 Maximum Entropy 91.65% 92.23% 91.94 Memory Based (MBL) 91.05% 92.03% 91.54 Markov Models 90.63% 89.65% 90.14 Rules 86.24% 88.25% 87.23 Rules 88.82% 82.91% 85.76 EDVHOLQH 72.58% 82.14% 77.07 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 54 Shallow Parsing VI * Best results for combined methods and specially a dynamic programming combination of SVMs * With these results ­ It is possible to accurately detect noun phrases for indexing in text classification systems ­ But probably more precision is required for Information extraction 28 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 55 Information Extraction I * The goal of Information extraction (IE) is transform text into a structured format (e.g. database records) according to its content ­ E.g. Heterogeneous researchers homepages are transformed into database records containing name, position, institution, research interests, projects, etc ­ E.g. Terrorism news articles are transformed into records including kind of incident, place, date, instigator, personal damages, etc Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 56 Information Extraction II * A key application is feeding other text and mining processes (see e.g. [Nahm & Mooney 02] about the project DISCOTEX) * Also structured databases are given to analysts to support their work, e.g. finding trends and forecasting according to them * Introduction regarding web content [Eikvill 99] * See [Cowie & Lehnert 96, Cunningham 99] 29 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 57 Information Extraction III * Techniques range from knowledge poor to rich, with obvious increasing domain dependence and effectiveness * Approaches depend on the structure of text ­ Free text with fully grammatical sentences allow natural language processing techniques with the induction of patters based on syntactic and semantic analysis ­ ... Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 58 Information Extraction IV ­ ... ­ Structured text follows a predefined and structured format that leads to delimiter based patterns ­ Semi-structured text is telegraphic and ungrammatical, thus a combination of techniques that employ several sources of information are used 30 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 59 Information Extraction V * A popular approach is the automatic induction or manual derivation of ZUDSSHUV (see e.g. [Freitag & Kushmerick 00]) * A wrapper is a procedure for extracting a particular resource's content * Usually consists of a set of extraction rules and a rule engine * A wrapper is information source dependent Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 60 Information Extraction VI * Wrapper construction can be done through inductive learning, by reasoning about a sample of the resource's documents * Kushmerick et al. have identified several classes of wrappers which are UHDVRQDEO\ XVHIXO \HW HIILFLHQWO\ OHDUQDEOH * To assess usefulness, they measured the fraction of Internet resources that can be handled by their techniques and found that their system can learn wrappers for 70% of the surveyed sites 31 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 61 Summary * In this track we have presented a sample of tasks and techniques that ­ Are oriented to supervised learning from text ­ In the context of TM * In a "real" TM environment (in the sense by Hearst), the tasks are successively applied to texts, and combined with unsupervised tasks and techniques 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ /($51,1* :+(1 127 .12:,1* 1 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 1 Outline 1. Introduction 2. Document Clustering 3. Term Clustering 4. Document Summarization Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 2 Introduction * Unsupervised learning tasks ­ Unlabelled textual data ­ Limited concept quality metrics in dimensionality reduction ­ Unsupervised induction of text groups through data clustering algorithms ­ Coherence or task-dependent evaluation metrics * More strict TM tasks 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 3 Document Clustering I * Document Clustering (DC) is partitioning a set of documents into groups or clusters * Clusters should be computed to ­ Contain similar documents ­ Separate as much as possible different documents * For instance, if similarity between documents is defined to capture semantic relatedness, documents in a cluster should deal with the same topics, and topics in each cluster should be different Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 4 Document Clustering II * DC applications include ­ Exploratory text data analysis * In words by Manning & Schütze [99] ",W LV DOZD\V D PLVWDNH WR QRW ILUVW VSHQG VRPH WLPH JHWWLQJ D IHHO IRU ZKDW WKH GDWD DW KDQG ORRN OLNH" (p. 497) ­ Pre-processing for other tasks, e.g. * In [Karypis & Han 00], to detect the main semantic dimensions of the text collection and to support a kind of &RQFHSW ,QGH[LQJ) * In [Hatzivassiloglou et al. 01] for text summarization 3 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 5 Document Clustering III * But the main application is supporting a variety of Information Access tasks ­ Guiding the organization of a document collection (e.g. [Sahami 98]) * Progressively clustering groups of documents allow to build a topic hierarchy ­ Supporting browsing and interactive retrieval (e.g. [Cutting et al. 92b, Baldonado & Winograd 97, Wu et al. 01]), now some search engines (Vivisimo) * Grouping retrieved documents to allow a faster relevant documents selection process Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 6 Document Clustering IV * As other text processing tasks, DC has several steps ­ Document representation ­ Dimensionality reduction ­ Applying a clustering algorithm ­ Evaluating the effectiveness of the process * Basically, we can apply the same methods that we use in data clustering (see the survey by Jain et al. [99]) 4 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 7 Document Clustering V * Document representation ­ Again, documents are represented as concept weight vectors ­ Concepts are usually words filtered according stop lists and stemmed ­ Weighting approaches include binary, WI, WILGI, etc Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 8 Document Clustering VI * Dimensionality reduction ­ Two main approaches, depending on the knowledge used * If we have not knowledge, distributional approaches are often applied: the Zipf's law ­ Very frequent and very unusual terms are filtered out ­ They have less discriminative power (see [Salton & McGill 83; Salton 89] * If we have knowledge (e.g. we have query concepts in interactive retrieval), a number of heuristics are used (see e.g. [Rüger & Gauch 00]) 5 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 9 Document Clustering VII * Dimensionality reduction ­ The Zipf's law and discriminative power * Most discriminative concepts have low to medium frequency frequency rank min max discriminative power Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 10 Document Clustering VIII * Clustering algorithms ­ Include a wide sample of the available in all required conditions * Hierarchical versus flat * Hard versus soft * Several semantic similarity functions ­ Perhaps the most often applied method is a kind of Hierarchical Agglomerative Clustering (HAC) method, but sometimes Expectation-Maximization (AutoClass) and Self-Organizing Maps 6 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 11 Document Clustering IX * HAC (as in [Manning & Schütze 99]) ­ Starts with a cluster per document ­ In each iteration, the closest pair of clusters are merged ­ It depends on the way similarity between documents and between clusters is defined * Typical inter document similarity metrics are cosine, Euclidean, etc * Similarity between clusters can be measured by single link, complete link, group-average, etc methods Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 12 Document Clustering X * EM (as in [Manning & Schütze 99]) ­ Can be seen as a way of estimating the hidden parameters of a model * Given some data X, and a model M with parameters, we want to estimate P(X|M()) and to find the model that best fits the data (maximizes the likelihood of the data) ­ Beginning with an approximation of , iterates two steps * In the expectation step, the parameters of the model are estimated and interpreted as cluster membership probabilities * In the maximization step, the most likely parameters are estimated given the cluster membership probabilities 7 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 13 Document Clustering XI * DC quality metrics ­ Sometimes, the results are compared to a manual classification taken as golden standard leading to accuracy based metrics (entropy, F, Mutual Information, etc) (see e.g. [Zhao & Karypis 02, Vaithyanathan & Dom 99]) ­ With no information of class labels, metrics like overall similarity (cohesiveness) are used (e.g. [Steinbach et al. 00]) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 14 Document Clustering XII * DC quality metrics ­ Also indirect evaluation (in the context of a second task) is possible ­ For instance, the increase of retrieval effectiveness in a text retrieval experiment including * Direct retrieval effectiveness metrics (recall, precision, etc) [Leuski 01] * Time to find the information [Hatzivassiloglou et al. 01] 8 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 15 Term Clustering I * This is an umbrella that includes a number of tasks and techniques, e.g ­ Automatic thesaurus construction ­ Latent semantic indexing * The basic idea is to work at the word level to develop models usable in indexing for other document level tasks Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 16 Term Clustering II * Automatic Thesaurus Construction (ATC) ­ A thesaurus is (traditionally) a dictionary of synonyms, but the concept has evolved to include semantic relations between concepts * Class #764 of an engineering related thesaurus (refusal) refusal declining non-compliance rejection denial ­ A text processing oriented thesaurus contains * Concepts that contain words and multi-word expressions * Relations like is-a and has-a ­ In fact, semantic nets 9 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 17 Term Clustering III * Automatic Thesaurus Construction ­ Research in ATC begins with IR (50's) (see consolidated references as Salton's) ­ The problem of lack of statistics for individual but related words is addressed by grouping semantically related words into classes ­ The classes in the thesaurus are after used for better representing document contents Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 18 Term Clustering IV * Automatic Thesaurus Construction ­ It is possible to use the same techniques as those for document clustering ­ Traditional IR techniques are based on * The hypothesis that related words occur in similar contexts * So the similarity between words is computed through the similarity between documents in which occur 10 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 19 Term Clustering V * Automatic Thesaurus Construction ­ So we work on a document x concept matrix ­ We compute similarity between concepts as similarity between the concept columns ­ Similarity between clusters is computed also using single link, complete link, group-average, etc approaches Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 20 Term Clustering VI * Automatic Thesaurus Construction ­ Similarity between concepts may be computed through the cosine formula * Where wdjk y wdjl are the weights of the kth and lth concepts in the jth document of the collection == = = P M MO P M MN P M MOMN ZGZG ZGZG ONVLP 1 2 1 2 1 ),( 11 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 21 Term Clustering VII * Automatic Thesaurus Construction ­ Again a variety of clustering methods can be used, being HAC the most frequent ­ While intuition supports that using semantic classes should lead to better representation, experience in IR is negative [Salton & McGill 83, Salton 89] ­ For instance, a relevance feedback iteration can be 10 times better than it Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 22 Term Clustering VIII * Latent Semantic Indexing ­ It has been presented as a way to capture the main semantic dimensions in a text collection, avoiding synonymy and polysemy problems ­ Exploits co-occurrence information between concepts to derive a text representation based on new, less dimensions ­ Can be seen as an effective dimensionality reduction method 12 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 23 Term Clustering IX * Latent Semantic Indexing ­ Conceived for IR [Deerwester et al. 90, Berry et al. 95] ­ E.g. applied to * TC [Dumais & Nielssen 92] * Filtering [Dumais 95] * Cross-Language IR [Dumais et al. 97] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 24 Term Clustering X * Latent Semantic Indexing ­ E.g. (borrowed from [Manning & Schütze 99]) Let A a concept x doc matrix for a text collection = 101000truck 011001car 000011moon 000010astronaut 000101monautcos dddddd A 654321 13 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 25 Term Clustering XI * Latent Semantic Indexing ­ E.g. (borrowed from [Manning & Schütze 99]) ­ Given A, it can be observed that * Cosmonaut & astronaut are synonyms and never cooccur, but they do with moon * sim(d2,d3)=0 but they contain synonym concepts * sim(d5,d6)=0 but they contain synonym concepts ­ Of course, there are two main semantic dimensions in the data (astronomy, road) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 26 Term Clustering XII * Latent Semantic Indexing ­ E.g. (borrowed from [Manning & Schütze 99]) ­ After LSI has applied, reducing to 2 dimensions d1 d2 d3 d4 d5 d6 14 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 27 Term Clustering XIII * Latent Semantic Indexing ­ The basic idea is mapping a high-dimensional space into a low-dimensional one ­ Iteratively choosing dimensions corresponding to the axes of greater variation ­ Co-occurring concepts are mapped onto the same dimension ­ A method called Singular Value Decomposition for the analysis of co-occurrence patterns is the core Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 28 Document Summarization I * Document Summarization is the task of abstracting key content from one or more information sources * It can be seen as a classification (knowledge poor) or understanding (knowledge rich) task * Interesting overview in [Hahn & Mani 00] 15 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 29 Document Summarization II * Mostly applied to ease information access ­ E.g. Most useful keywords are extracted from a set of documents (e.g. a cluster) to describe it ­ E.g. Documents in a collection are abstracted to avoid reading the full content ­ E.g. Documents retrieved from search are summarized to allow the user a faster identification of those relevant to the query Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 30 Document Summarization III * We will classify approaches by size of the text unit used in the summary ­ Keyword summaries ­ Sentence summaries * Although many classification based techniques for sentence summaries can be applied to keyword summaries 16 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 31 Document Summarization IV * Keyword summaries ­ Abstracting is equivalent to detect most informative keywords or multi word expressions in a (set of) document(s) ­ For one document, it can be as simple as selecting the higher weight concepts in it (according to WILGI) ­ In a interactive retrieval setting * should make the document as different as possible to other non related documents * should explain why the document has been retrieved Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 32 Document Summarization V * Sentence summaries (according to [Hahn & Mani 00]) * Two taxonomies ­ Regarding function * Indicative * Informative * Critical ­ Regarding target user * Generic * Used-focused 17 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 33 Document Summarization VI * Sentence summaries [Hahn & Mani 00] ­ The basic steps are * Analysing the source text * Determining its salient points * Synthesizing an appropriate output ­ We will focus on how knowledge poor (classification based) summarizing systems address these steps Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 34 Document Summarization VII * Sentence summaries [Hahn & Mani 00] ­ Text units are sentences ­ A linear model of salience is often applied using a set of features (heuristics) * Location in the source text * Appearance of cue phrases * Statistical significance * Additional information Weight(U) = Location(U) + CuePhrase(U) + StatTerm(U) + AddTerm(U) 18 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 35 Document Summarization VIII * Sentence summaries [Hahn & Mani 00] ­ Location criterion * Weight sentences according to the part of the paragraph or document they occur in ­ Most news stories begin with a small summary ­ Sentences in the introduction and conclusions of research papers are very likely to occur in the summary ­ Cue Phrase criterion * Lexical or phrasal summaries as "in conclusion", "in this paper", etc * The approach suggest to overweight the sentences in which they occur Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 36 Document Summarization IX * Sentence summaries [Hahn & Mani 00] ­ Statistical salience * Well known IR heuristic weights as WILGI applied to select those sentences in which the concept occur ­ Additional term * Depending on the application, we can ­ Promote sentences that include query concepts (retrieval => query biased summaries) ­ Promote sentences that include user profile concepts (filtering => used adapted summaries) 19 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 37 Document Summarization X * Sentence summaries ­ In [Kupiec et al. 95] * A supervised learning approach has been devised (and applied to the paper itself) * It has been shown that location combined with cue phrases is a very powerful method ­ In [Maa et al. 99] * Query biased summaries demonstrate their informativeness in a relevance feedback process Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 38 Document Summarization XI Process in a supervised learning based summarizer [Hahn & Mani 00] 20 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 39 Document Summarization XII * Sentence summaries ­ Evaluation * Intrinsic evaluation using human judgements (gold standard) * Extrinsic evaluation ­ The summarizer is as good as it contributes to a task in which is applied ­ For instance, the informativeness of the summary can be measured in terms of the accuracy of a summary based retrieval in comparison with a full-document retrieval Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 40 Document Summarization XIII * A note about knowledge-rich summarization ­ A very common approach is to fill a template with facts in the text and after producing a canned-text summary * Filling the template is a IE task * Domain dependent ­ But more complex systems have been devised, including the attempt to capture meaning and appropriate output planning (e.g. SUMMARIST [Hovy & Lin 99]) 21 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 41 Summary * In this track we have presented a sample of tasks and techniques that ­ Are mostly oriented to unsupervised learning from text in the context of TM ­ Specially clustering-based problems are closer to real TM * Again, in a "real" TM environment (in the sense by Hearst), the tasks are successively applied to texts, and combined with supervised tasks and techniques 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ 722/6 1 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 1 Outline 1. Introduction 2. Tools for Text Mining 3. IBM Intelligent Miner for Text 4. University of Sheffield General Architecture for Language Engineering (GATE) 5. University of Waikato Environment for Knowledge Analysis (WEKA) NOTE: all tools are a trade mark of their respective companies Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 2 Introduction * The goals of this track are ­ Briefly reviewing the TM tools market state of the art with special attention to consolidated systems ­ Give an overview of a commercial tool and a research tool ­ Introduce the tool to be used in this tutorial, WEKA 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 3 Tools for Text Mining I * Many KDD tools can be used for TM, provided you add text processing functions * But the increasing market of specialized commercial TM tools shows the interest of the topic * A review in [Tan 99] discuss only 11 products, but The Data Warehousing Information Center lists 100 tools and vendors (see the web page) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 4 Tools for Text Mining II * Most tools provide a subset of the following functionalities ­ Information Extraction ­ Text Retrieval ­ Text Categorization ­ Text Clustering ­ Text Summarization ­ Visualization ­ And a sample of word level techniques (e.g. POS Tagging) * But nearly none is a "real" TM tool (in Hearst's sense) 3 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 5 IBM Intelligent Miner for Text I * A very representative example of state of the art commercial TM environment is IBM Intelligent Miner for Text (IMT) * Provides support for most of the listed tasks plus some other useful functionalities * It is sketched in [Tkach 98] Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 6 IBM Intelligent Miner for Text II * IBM IMT provides support for ­ Language recognition ­ Named entity recognition ­ Document clustering (HAC) ­ Text categorization (rule induction, kNN) ­ Text retrieval ­ Text summarization ­ Some support for other languages than English * And other helpful elements as web spiders 4 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 7 IBM Intelligent Miner for Text III * IBM IMT has been conceived as TM library, oriented to user design of ­ Business applications, e.g. * Customer e-mail processing * The journalist's workstation ­ Intranet/Internet applications, e.g. * "Show me more like this" powered searches * Topic dependent searches Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 8 IBM Intelligent Miner for Text IV Clustering example (borrowed from IBM IMT Getting Started) 5 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 9 IBM Intelligent Miner for Text V Feature extraction (borrowed from IBM IMT Getting Started) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 10 Sheffield's GATE I * The University of Sheffield General Architecture for Text Engineering (GATE) is architecture, framework and development environment for /DQJXDJH (QJLQHHULQJ (LE) * GATE includes more than 700 Java classes (more than 15 Mb of byte code) and it is open source software (under GNU's GPL) * See [Cunningham et al. 02] for an overview 6 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 11 Sheffield's GATE II * It has been designed to ­ Cleanly separate a number text processing tasks ­ Allow automatic measurement of performance ­ Reduce integration overheads ­ Provide a set of basic language processing components to extended or replaced * As an architecture ­ It defines the organization of a LE system and the assignment of responsibilities to different components ­ It ensures the component interactions satisfy the system requirements Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 12 Sheffield's GATE III * As a framework ­ It provides a reusable design for LE systems, and a set of prefabricated software blocks to be used, extended or customised to meet specific needs * As development environment ­ It helps its users to minimise the time spent to build of modify LE systems, e.g with the debugging mechanism 7 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 13 Sheffield's GATE IV * GATE is best understood through an application * ANNIE (A Nearly-New Information Extraction system) is a IE tool designed to extract relevant information about people from their home pages * ANNIE components form a pipeline (as text processing is data-intensive, in cascade) ­ See GATE documentation Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 14 Sheffield's GATE V ANNIE modules (borrowed from GATE User Manual) 8 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 15 Sheffield's GATE VI * What GATE lacks of ­ Learning machinery!!! * But this role can be played by other KDD tools, being WEKA a very suitable one Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 16 WEKA I * WEKA is the Waikato Environment for Knowledge Analysis * It is a fully reusable set of KDD tools including feature selection, ML algorithms, evaluation and visualization * It consists of around 360 Java classes (more than 1Mb of byte code) and it is open source software (under GNU's GPL) * See [Witten & Frank 99] for details 9 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 17 WEKA II * It is important to note that WEKA is not text oriented * Currently, you can ­ Either program yourself the text processing tools required ­ Use other packages (e.g. Smart, etc) for text and manage the integration ­ Join and follow up WETA Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 18 WEKA III * WETA (Waikato Environment for Text Analysis) is an initiative framed in the 2SHQ1/3 SURMHFW * Its goal is to develop a highly scalable solution for text analysis based on machine learning algorithms contained in WEKA * Still far from it 10 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 19 WEKA IV * WEKA usage modes ­ For testing different approaches to a learning problem (command-line, GUI) ­ For developing applications that make use of learning ­ For researching in and developing new algorithms Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 20 WEKA V * WEKA for testing learning approaches ­ Most classes provide a main method for command-line usage (designed for testing) ­ A relatively sophisticated GUI is provides, with three options * Simple CLI ­ acts as a command line interface * Explorer ­ designed for processing, learning, evaluation and visualization * Experimenter ­ designed for distributing intensive processing experiments 11 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 21 WEKA VI * WEKA Explorer ­ Tabs for processing, classifying, clustering, compute associations, perform attribute selection, and visualization ­ A typical operation procedure involves * Loading data in Preprocess * Visualizing data in Visualize * And iteratively ­ Perform attribute selection ­ Test (several) learning algorithm(s) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 22 WEKA VII Preprocessing tab in WEKA Explorer 12 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 23 WEKA VIII Classification tab in WEKA Explorer Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 24 WEKA IX Visualizing a threshold curve in WEKA Explorer 13 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 25 WEKA X Visualizing a decision tree in WEKA Explorer Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 26 WEKA XI * WEKA for developing learning based applications ­ It provides a comprehensive API that allows the development of applications ­ E.g. The tutorial example implements a simple email message recommendation program that learns (kNN) your interests regarding a number of keywords 14 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 27 Summary * A wide range of "shallow" TM commercial tools, covering many useful tasks and oriented to enterprise environments * Some research but still useful open source tools * We believe it is possible to develop high quality, advanced TM applications with the GATE + WEKA combination 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ '(7(&7,1* 32512*5$3+< 1 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 1 Outline 1. Motivation 2. Current Technology 3. The POESIA Project 4. POESIA Text Filter by the UEM Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 2 Motivation I * Current social interest in filtering & blocking solutions for a range of environments * E.g. Kids surfing for fun at the school or library * E.G. Employees surfing for fun at work * The common feature is misusing Internet in a place in which it should be used for a different purpose 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 3 Motivation II * There is is a market and an industry ­ Internet users most frequent search is sex (by February 2001) ­ The 8,5% of search engine queries deal with sex and pornography (2001) ­ 27.5 million of U.S citizens visited pornographic websites in January 2002 ­ U.S. citizens spent $220 million on 2001 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 4 Motivation III * But also there is a problem * E.g. for kids ­ One of five 10 to 17 years youngsters was asked for sex in 2000 ­ One of four had access to unwanted explicit sex stuff * E.g. for companies ­ Internet abuse at the workplace produced $1 billion loss in 2001 3 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 5 Motivation IV * On kids safe access to the Internet, there is an ongoing activity by a number of government agencies, including e.g. ­ The Safer Internet Action Plan by the European Commission ­ The ITAS by the US National Research Council ­ NetAlert and the Australian Broadcasting Authority * On employees Internet misuse at the workplace, there is e.g. an interesting monographic issue in Communications of the ACM (January 02) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 6 Current Technology I * In recent years, some reports about Internet filtering technology and effectiveness have been published, e.g. ­ NetProtect Report on Filtering Techniques and Approaches ­ CSIRO Report on Effectiveness of Internet Filtering Software Products * There is an increasing number of commercial and research solutions in the market 4 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 7 Current Technology II * Current commercial products include some the following techniques ­ Black and white lists ­ Self and third-party labelling (ICRA, PICS) ­ Keyword based text processing ­ Image processing by skin detection * Most techniques have been found quite ineffective in isolation and are rarely used in combination Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 8 Current Technology III * To our knowledge, there is only one research paper dealing with text based pornography detection (project FILTERIX) [Chandrinos et al. 00] * They address the problem as TC with ­ Text representation as binary weight vectors ­ Information Gain for feature selection ­ Naive Bayes learning for classification * With promising effectiveness results 5 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 9 The POESIA Project I * The work described hereafter is a part of the POESIA project * POESIA stands for Public Open-source Environment for a Safer Internet Access * POESIA aims to develop, test, evaluate and promote a fully open-source, extensible, state of the art, filtering and caching software solution, targeted for situations where browsing and other Internet activities are undertaken, e.g. classrooms * See http://www.poesia-filter.org Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 10 The POESIA Project II * POESIA will ­ Cover at least Web and incoming email channels ­ Filter at least the pornographic and offensive speech domains ­ Target English, Spanish and Italian languages * Starting on Feb. 2001, and 2 years long * Partly funded by the EC Safer Internet Action Plan 6 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 11 The POESIA Project III * Partners at POESIA ­ Istituto di Linguistica Computazionale (Italy) ­ Commissariat l'Energie Atomique (France) ­ Ecole Nouvelle d'Ingénieurs en Communication (France) ­ M.E.T.A. S.r.l. (Italy) ­ Universidad Europea de Madrid CEES (Spain) ­ University of Sheffield (UK) ­ Fundació Catalana per a la Recerca (Spain) ­ PIXEL Associazione (Italy) ­ Liverpool Hope University College (UK) ­ Telefónica Investigación y Desarrollo (Spain) Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 12 The POESIA Project IV * Technologic approach ­ Combination of a number of techniques including * Label detection * Sophisticated text analysis * Sophisticated image processing * Script code analysis * Effectiveness is got by applying state-of-theart research approaches * Efficiency is got through a two level schema and caching facilities 7 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 13 The POESIA Project V * We focus on the text processing approaches ­ Under development by Sheffield, ILC and UEM ­ Two-stage architecture 1. A simple (`lite') filtering agent which makes only light use of NLP techniques, and can rapidly process large text volumes 2. A sophisticated (`heavy') filtering agent which makes heavier use of NLP resources and techniques to filter only those documents that are left uncategorized by the first agent ­ Addressed as a TC task Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 14 The POESIA Project VI * The most sensible approach for the lite text filter, given current state of the art, is ­ Text representation based on binary or WILGI weight vectors ­ Feature selection with Information Gain, 2, etc ­ Learning with Support Vector Machines and possibly some cost sensitive method (e.g. MetaCost) 8 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 15 The POESIA Project VII * Current Spanish text filtering prototype ­ Techniques * Binary weight vectors * IG * SVM ­ Effectiveness * English accuracy 92% * Spanish accuracy 88% Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 16 The POESIA Project VIII * Current Spanish text filtering prototype ­ Technology * 26 Java classes * 2600 code lines * Reusing ­ WEKA ­ HTMLParser ­ Muffin proxy filter demonstration 9 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 17 The POESIA Project IX * Current Spanish text filtering prototype ­ Design * A HTML parsing package * A binary index package * Two Muffin filter classes ­ Operation * Learning step ­ A set of manually classified web pages are indexed and a WEKA SVM classifier trained on them * Classification step ­ The demanded web page is processed and classified on-the-fly 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ '(7(&7,1* 63$0 1 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 1 Outline 1. Motivation 2. State of the Art 3. Problem Description 4. Evaluation Framework 5. Comparison of Approaches 6. Results and Conclusions 7. Notes about implementation Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 2 Motivation * Spam email is more properly, Unsolicited Bulk Email (UBE) * It has been producing a considerable damage to ­ Internet Service Providers ­ Internet users ­ and the whole Internet backbone * For instance, Internet subscribers worldwide are wasting an estimated 10 billion euro a year just in connection costs due to UBE 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 3 State of the Art I * Three kinds of proposals to address UBE [Cranor & LaMacchia 98, Hoffman & Crocker 98] ­ Economic * Charging sending email ­ Regularory * Law definition and enforcement ­ Technical * Filtering mechanisms Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 4 State of the Art II * Technical approaches ­ Channels [Hall 99] ­ Aliasing [Gabber et al. 99] ­ Filtering * Black & white lists * Bulk message detection * Content-based detection ­ Manual filters (e.g. the one operated by BrightMail) ­ Machine learned classifiers (see e.g. [Sahami et al. 98, Gómez et al. 00, Gómez et al. 02] and the bibliography at http://liinwww.ira.uka.de/bibliography/Ai/MLSpamBibliography.html ) 3 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 5 State of the Art III * Very good reported results for manual filters in a study by ETesting Labs [Etesting Labs 01] * Brightmail approach seems very effective ­ Able to catch 93.9% of UBE without missclassifying any legitimate message ­ Based on millions of email addresses receiving UBE and a team of experts manually codifying rules on-the-fly Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 6 Problem Description I * UBE detection is a TC problem ­ Two classes (UBE and legitimate email) ­ It relatively easy to * Represent messages as vectors of concept weights * Perform some feature selection * Learn a classifier * But evaluation is not so simple because it is a problem in which missclassification costs and class distribution are not symmetric 4 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 7 Problem Description II * It is clear that users prefer dealing with more UBE to missing legitimate email * But the preference is not rated anywhere, i.e. we do not know the relative costs of both kinds of mistakes ­ E.g. It is one hundred times worse missing a legitimate email than receiving a UBE? * Even worse, these costs may vary Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 8 Problem Description III * So we need 1. A method for evaluating classification accuracy that is independent of class and cost distributions 2. To consider cost-sensitive learning methods (e.g. stratification and weighting, threshold variation, MetaCost, BoostCost, etc) 5 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 9 Evaluation Framework I * We use the Receiver Operating Characteristic Convex Hull (ROCCH) method [Drummond & Holte 00, Provost & Fawcett 97, 01] * ROC analysis allows a visual comparison of the performance of a set of ML algorithms, regardless of the class and cost conditions * With the study of the Convex Hull, we can detect the best approach for the required class and cost distributions Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 10 Evaluation Framework II * A ROC curve for an algorithm is produced by ­ Learning clasffiers for a variety of conditions and linking the set of (false positive, true positive) points in a 2D graph ­ Or going through the rank of classified instances and obtaining the set of (fp,tp) points * We follow the first approach ­ It is rather time consuming but you can store the learned classifiers and use the best for some aplication environment 6 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 11 Evaluation Framework III * Sketch of evaluation method 1. For each ML algorithm, obtain a ROC curve and plot it (or only its convex hull) on the ROC space 2. Find the convex hull of the set of ROC curves previously plotted 3. Find the range of slopes for which each ROC curve lies on the convex hull 4. In case that target conditions are known, compute the corresponding slope value and output the best algorithm. In other case, output all ranges and best local algorithms or classiffers Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 12 Evaluation Framework IV * Given a class and cost distribution a slope can be computed as * Being c(Y,n) and c(N,y) the cost of a false positivive and a false negative * And P(y) and P(n) the probabilities of positive and negative classes )p(P)p,N(c )n(P)n,Y(c m = 7 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 13 Comparison of Approaches I * We have compared a number of TC approaches for detecting UBE (see [Gómez 02]) with ­ Text representation as binary and tf.idf weight vectors ­ Feature selection with IG, reducing the original concept space to 1% Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 14 Comparison of Approaches II * We have tested a number of ML algorithms ­ C4.5 ­ Naive Bayes ­ The rule learner PART ­ Support Vector Machines ­ The Rocchio algorithm 8 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 15 Comparison of Approaches III * We have tested a number of ML methods for making algorithms cost-sensitive ­ The Threshold method ­ The Weighting method (equivaleny to stratication by oversampling) ­ The MetaCost method Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 16 Results and Conclusions I ROCCH curve for the comparison of best classifiers and cost-sensitive methods 9 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 17 Results and Conclusions II Optimality ranges for the best classifiers and cost-sensitive methods Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 18 Results and Conclusions III * Regarding cost-sensitive methods ­ No one is clearly superior ­ Instance Weighting is the most frequent winner * Regarding learning algorithms ­ No one is clearly superior ­ SVM is the most frequent winner 10 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 19 Results and Conclusions IV * There are three scenarios considered important in the literature, corresponding to cost distributions in which a false positive is 1, 9 and 999 worse than a false negative Out results for the scenarios Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 20 Results and Conclusions V * Our results are better than those reported by others * More interestingly, for extreme conditions, we are close to real world manual performance ­ Brightmail detects 93.1% UBE without false positives ­ We get 91.7% UBE without false positives 11 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 21 Notes about Implementation * We used a the Ling-spam test collection * We processed the messages with the retrieval engine Smart to get the representation * But the bulk of the work has been done with WEKA * With our work in POESIA, we can used the implemented indexing tools to allow interacting with Smart 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ &+$//(1*(6 1 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 1 The challenge: supporting real TM * There are a number of TM related tools ­ They provide support for advanced text analysis tasks ­ They allow to discover new knowledge to the user but not the author * But we defined real TM as discovering absolutely new knowledge Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 2 The Challenge: supporting real TM * E.g. The example cited by Heast [99], describing Swanson's work ­ Given * medical titles and abstracts * a problem (incurable rare disease) * some medical expertise ­ find causal links among titles * symptoms * drugs * results 2 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 3 The Challenge: supporting real TM * E.g. The example cited by Heast [99] ­ Problem: Migraine headaches (M) * stress associated with M * stress leads to loss of magnesium * calcium channel blockers prevent some M * magnesium is a natural calcium channel blocker * spreading cortical depression (SCD) implicated in M * high levels of magnesium inhibit SCD * M patients have high platelet aggregability * magnesium can suppress platelet aggregability ­ All extracted from medical journal titles Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 4 The Challenge: supporting real TM * E.g. The example cited by Heast [99] ­ These clues suggest that magnesium deficiency may play a role in some kinds of migraine headache ­ The hypothesis which did not exist in the literature at the time Swanson found these links ­ The hypothesis must of course be confirmed experimentally ­ The process to derive the hypothesis was not automatic 3 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 5 The Challenge: supporting real TM * So what we need is tools to strongly support this kind of knowledge discovery * Hearst describes a project called LINDI focused on developing this kind of tool for finding functions of genes * The idea behind the system is mixed-initiative interaction ­ User applies tools to help explore the hypothesis space ­ System runs suites of algorithms to help explore the space, suggest directions Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 6 The Challenge: supporting real TM * The system has three main parts ­ UI for building/using strategies ­ Backend for interfacing with various databases and translating different formats ­ Content analysis/machine learning for figuring out good hypotheses/throwing out bad ones 4 Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 7 The Challenge: supporting real TM * A point is that we have open-source tools to develop this kind of systems, in a three layered architecture ­ GATE may act as backend ­ WEKA+GATE may act the learnening-based middleware ­ UI and integration are required Text Mining and Internet Content Filtering, ECML/PKDD Tutorial, August 19th, 2002 8 The Challenge: supporting real TM * But still we need further investigation in several fields [Tan 99], specially ­ Multilinguality ­ Domain knowledge integration * And of course all the challenges in KDD 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ 5()(5(1&(6 1 7H[W 0LQLQJ DQG ,QWHUQHW &RQWHQW )LOWHULQJ 5HIHUHQFHV IRU WKH (&0/3.'' 7XWRULDO $XJXVW WK José María Gómez Hidalgo Departamento de Inteligencia Artificial Universidad Europea CEES jmgomez@dinar.esi.uem.es http://www.esi.uem.es/~jmgomez/ [Abney 91] Steven Abney. Parsing By Chunks. In: Robert Berwick, Steven Abney and Carol Tenny (eds.), Principle-Based Parsing. Kluwer Academic Publishers, Dordrecht. 1991. [Agirre & Rigau96] Agirre E. and Rigau G., Word Sense Disambiguation using Conceptual Density. Proceedings of 15th International Conference on Computational Linguistics, COLING'96. Copenhagen, Denmark, 1996. [Allan 96] J. Allan. Incremental relevance feedback for information filtering. In Proc. ACM SIGIR Conf., Zurich, Switzerland, August 1996. [Attardi et al. 99] Giuseppe Attardi, Antonio Gullí and Fabrizio Sebastiani, Automatic Web Page Categorization by Link and Context Analysis. In Proceedings of THAI-99, 1st European Symposium on Telematics, Hypermedia and Artificial Intelligence, pp. 105-119, 1999. [Baldonado & Winograd 97] Baldonado, M.Q.W., and Winograd, T. SenseMaker: An Information-Exploration Interface Supporting the Contextual Evaluation of a User's Interest, In proceedings of CHI '97, Atlanta, GA [Benkhalifa et al.01] Mohammed Benkhalifa, Abdelhak Mouradi and Houssaine Bouyakhf. Integrating WordNet knowledge to supplement training data in semisupervised agglomerative hierarchical clustering for text categorization. International Journal of Intelligent Systems, 16(8), pp. 929-947, 2001. [Berry et al. 95] Berry, M. W., Dumais, S. T., and O'Brien, G. W. (1995). "Using linear algebra for intelligent information retrieval." SIAM Review, 37(4), 1995, 573-595. [Bloerdon et al.96] Bloedorn, Eric, Inderjeet Mani, and T. Richard MacMillan. Representational Issues in Machine Learning of User Profiles, 1996 AAAI Spring Symposium on Machine Learning in Information Access, 1996. [Brill95] Eric Brill, Transformation-Based Error-Driven Learning and Natural Language Processing: A Case Study in Part of Speech Tagging, Computational Linguistics, Vol 21, No 4, 1995. [Buenaga et al. 00] de Buenaga Rodríguez, M., Gómez Hidalgo, J.M., Díaz Agudo, B. Using Wordnet to Complement Training Information in Text Categorization, in Nicolov, N. and Mitkov, R. (eds) Recent Advances in Natural Language Processing II: Selected Papers from RANLP'97, Current Issues in Linguistic Theory (CILT), vol. 189, John Benjamins: Amsterdam/Philadelphia, 2000. 2 [Caropreso et al.01] Maria Fernanda Caropreso, Stan Matwin and Fabrizio Sebastiani. A learner-independent evaluation of the usefulness of statistical phrases for automated text categorization. Text Databases and Document Management: Theory and Practice, pp. 78-102, Idea Group Publishing, 2001. [Chakrabarti et al.98] Soumen Chakrabarti, Byron E. Dom, and Piotr Indyk. Enhanced hypertext categorization using hyperlinks. In Proceedings of SIGMOD-98, ACM International Conference on Management of Data, pp. 307-318, ACM Press, New York, US, 1998. [Chandrinos et al. 00] K.V. Chandrinos, I. Androutsopoulos, G. Paliouras and C.D. Spyropoulos, "Automatic Web Rating: Filtering Obscene Content on the Web". In Borbinha, J. and Baker, T. (Eds.), Proceedings of the 4th European Conference on Research and Advanced Technology for Digital Libraries (ECDL 2000), Lisbon, Portugal, pp. 403-406, 2000. [Church 88] Church, K. (1988) "A Stochastic Parts Program and Noun Phrase Parser for Unrestricted Text," Second Conference on Applied Natural Language Processing, Austin, Texas, pp. 136-143. [Cowie & Lehnert 96] J. Cowie and W. Lehnert. Information Extraction. Communications of the ACM, 39(1):80--91, 1996. [Cranor & LaMacchia 98] Lorrie F. Cranor and Brian A. LaMacchia. Spam! Comm. of the ACM, 41(8), 1998. [Cunningham 99] H. Cunningham. Information Extraction: a User Guide (revised version). Department of Computer Science, University of Sheffield, May, 1999. [Cunningham et al. 02] H. Cunningham. GATE, a General Architecture for Text Engineering. Computing and the Humanities, Vol. 36, pp. 223-254, 2002. [Cutting et al. 92a] D. Cutting, J. Kupiec, J. Pedersen, and P. Sibun. A Practical Part-ofSpeech Tagger, Proceedings of the Third Conference on Applied Natural Language Processing, April 1992. [Cutting et al. 92b] D. Cutting , D. Karger, J. Pedersen, and J. Tukey. Scatter/Gather: A Cluster-based Approach to Browsing Large Document Collections, Proceedings of SIGIR'92, June 1992. [Deerwester et al.90] Deerwester, S., Dumais, S. T., Landauer, T. K., Furnas, G. W. and Harshman, R. A. (1990) - no figures, "Indexing by latent semantic analysis." Journal of the Society for Information Science, 41(6), 391-407. [DeRose 88] DeRose, Stephen J. 1988. Grammatical category disambiguation by statistical optimization. Computational Linguistics 14.1: 31-39. [Dörre et al. 99] J. Dorre, P. Gerstl, and R. Seiffert, 1999. Text Mining: Finding Nuggets in Mountains of Textual Data. In Proceedings of the 5th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Diego, California. [Drucker et al. 99] Harris Drucker, Vladimir Vapnik, and Dongui Wu. Automatic text categorization and its applications to text retrieval. IEEE Transactions on Neural Networks, 10(5):1048-1054, 1999. 3 [Drummond & Holte 00] Chris Drummond and Robert C. Holte. Explicitly representing expected cost: An alternative to ROC representation. In Proceedings of the Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 198-207, 2000. [Dumais & Nielssen 92] Dumais, S. T. and Nielsen, J. (1992), "Automating the assignment of submitted manuscripts to reviewers." In N. Belkin, P. Ingwersen, and A. M. Pejtersen (Eds.), SIGIR'92: Proceedings of the 15th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM Press, pp.233-244. [Dumais 95] Dumais, S. T. (1995), "Using LSI for information filtering: TREC-3 experiments." In: D. Harman (Ed.), The Third Text REtrieval Conference (TREC3) National Institute of Standards and Technology Special Publication , 1995. [Dumais et al. 97] Dumais, S. T., Letsche, T. A., Littman, M. L. and Landauer, T. K. (1997) "Automatic cross-language retrieval using Latent Semantic Indexing." In AAAI Spring Symposuim on Cross-Language Text and Speech Retrieval , March 1997. [Dumais et al. 98] Susan T. Dumais, John Platt, David Heckerman and Mehran Sahami. Inductive learning algorithms and representations for text categorization. Proceedings of CIKM-98, 7th ACM International Conference on Information and Knowledge Management, pp. 148-155, ACM Press, New York, US, 1998. [Eikvill 99] Line Eikvil: Information Extraction from World Wide Web - A Survey. Rapport Nr. 945, July, 1999. ISBN 82-539-0429-0 [EtestingLabs 01] eTesting Labs, Inc. Brightmail, inc. anti-spam service: Comparative performance test. Technical report, eTesting Labs, Inc., a Ziff Davis Media Inc. company, March 2001. Available at http://etestinglabs.com. [Fayyad et al. 96] Usama Fayyad, Gregory Piatetsky-Shapiro and Padhraic Smyth, 1996. Knowledge Discovery and Data Mining: Towards a Unifying Framework. In Proceedings of The Second Int. Conference on Knowledge Discovery and Data Mining, pages 82--88. [Feldman & Dagan 95] R. Feldman and I. Dagan, 1995. Knowledge Discovery in Textual Databases (KDT). In Proceedings of the 1st International Conference on Knowledge Discovery (KDD-95), pp. 112-117, Montreal. [Forsyth 99] Forsyth, R.S.1999. New directions in text categorization. In A. Gammerman Ed., Causal models and intelligent data management, pp.151-185. Heidelberg, DE:Springer. [Freitag & Kushmerick 00] Freitag, D. & Kushmerick, N. (2000). Boosted wrapper induction. AAAI-00 (Austin), pp. 577-583. [Gabber et al. 99] Eran Gabber, Phillip B. Gibbons, David M. Kristol, Yossi Matias, and Alain Mayer. Consistent, yet anonymous, Web access with LPWA. Communications of the ACM, 42(2):42-47, 1999. [Gale 92] William A. Gale, Kenneth Ward Church, David Yarowsky: Estimating Upper and Lower Bounds on the Performance of Word-Sense Disambiguation Programs. ACL 1992: 249-257. 4 [Ghani et al. 00] Rayid Ghani, Rosie Jones, Dunja Mladenic, Kamal Nigam and Sean Slattery, 2000. Data Mining on Symbolic Knowledge Extracted from the Web. In Proceedings of the Workshop on Text Mining [Gomez 02] Gómez Hidalgo, J.M. Evaluating Cost-Sensitive Unsolicited Bulk Email Categorization. ACM Symposium on Applied Computing, 2002. [Gómez et al. 00] Gómez Hidalgo, J.M., Maa López, M., Puertas Sanz, E. Combining Text and Heuristics for Cost-Sensitive Spam Filtering . Fourth Computational Natural Language Learning Workshop , CoNLL-2000, Lisbon, September 14, 2000. [Gómez et al. 02] Gómez Hidalgo, J.M, Puertas Sanz, E., Maa López, M. Evaluating Cost-Sensitive Unsolicited Bulk Email Categorization. 6th International Conference on the Statistical Analysis of Textual Data, Palais du Grand Large, St-Malo / France, March 13-15, 2002. [Gonzalo et al. 98] Gonzalo, J., F. Verdejo, I. Chugur and J. Cigarrán (1998) Indexing with WordNet synsets can improve text retrieval, in ACL/COLING Workshop on Usage of WordNet for Natural Language Processing. [Good et al. 99] Good, N., Schafer, J.B., Konstan, J., Borchers, A., Sarwar, B., Herlocker, J., and Riedl, J., Combining Collaborative Filtering with Personal Agents for Better Recommendations. Proceedings of the 1999 Conference of the American Association of Artifical Intelligence (AAAI-99). pp 439-446 [Grobelnik et al. 00] M. Grobelnik, D. Mladenic, and N. Milic-Frayling, 2000. Text Mining as Integration of Several Related Research Areas: Report on KDD'2000 Workshop on Text Mining. SIGKDD Explorations, December 2000, Volume 2, Issue 2, pp. 99-102. [Hahn & Mani 00] Hahn, Udo & Mani, Inderjeet (2000). The challenges of automatic summarization. In: Computer, 33 (11), pp. 29-36 [Hall 99][Hall 99] Robert J. Hall. A Countermeasure to Duplicate-detecting Anti-spam Techniques, AT&T Labs Research Report No. TR 99.9.1, 1999. [Hatzivassiloglou et al. 01] Hatzivassiloglou, V., J. Klavans, M. Holcombe, R. Barzilay, M.Y. Kan, K.R. McKeown. SimFinder: A Flexible Clustering Tool for Summarization. NAACL'01 Automatic Summarization Workshop. [Hearst & Karadi 97] Hearst, M. and Karadi, C. Cat-a-Cone: An Interactive Interface for Specifying Searches and Viewing Retrieval Results using a Large Category Hierarchy, Proceedings of the 20th Annual International ACM/SIGIR Conference , Philadelphia, PA, July 1997. [Hearst 95] Hearst, M. TileBars: Visualization of Term Distribution Information in Full Text Information Access, Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems (CHI), Denver, CO, 1995. [Hearst 99] Marti A. Hearst, 1999. Untangling Text Data Mining. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics, University of Maryland. [Hoffman & Crocker 98] Paul Homan and Dave Crocker. Unsolicited bulk email: Mechanisms for control. Technical Report Report UBE-SOL, IMCR-008, Internet Mail Consortium, 1998. 5 [Hovy & Lin 99] Hovy, E.H. and C-Y. Lin. 1999. Automated Text Summarization in SUMMARIST. In I. Mani and M. Maybury (eds), Advances in Automated Text Summarization. Cambridge: MIT Press, pp. 81-94. [Jain et al. 99] A.K. Jain, M.N. Murty and P.J. Flynn. Data Clustering: A Review. ACM Computing Surveys, Vol. 31, No. 3, September 1999 [Jelinek 85] Jelinek, F. (1985). Robust part-of-speech tagging using a hidden markov model. Technical report, IBM. [Junker & Abecker 97] Junker M. and Abecker A. (1997). Exploiting thesaurus knowledge in rule induction for text classification. In Milkov R., Nicolov N., and Nikolov N. editors, Proceedings of RANLP-97, 2nd International Conference on Recent Advances in Natural Language Processing, pages 202OE207, Tzigov Chark, BL. [Karypis & Han 00] George Karypis and Eui-Hong (Sam) Han. Concept Indexing: A Fast Dimensionality Reduction Algorithm with Applications to Document Retrieval & Categorization. In Proccedings of CIKM 2000 [Kessler et al. 97] Kessler,B. ,Nunberg,G. ,and Schütze,H.1997. Automatic detection of text genre. In Proceedings of ACL-97, 35th Annual Meeting of the Association for Computational Linguistics (Madrid, ES, 1997), pp.32 ­38. [Kodratoff 99] Kodratoff Y., 1999. Knowledge Discovery in Texts: A Definition, and Applications. Foundation of Intelligent Systems, Ras & Skowron (Eds.) LNAI 1609, Springer. [Kupiec et al. 95] Julian Kupiec, Jan Pedersen, and Francine Chen. A trainable document summarizer. In Proceedings of the 18th ACM/SIGIR Annual Conference on Research and Development in Information Retrieval, pages 68--73, Seattle, Washington, 1995. [Larkey 99] Larkey, L.S.1999. A patent search and classiFDWLRQ V\VWHP ,Q 3URFHHGLQJV of DL-99, 4th ACM Conference on Digital Libraries (Berkeley,US,1999),pp.179 ­187. [Lavrenko et al. 00] Victor Lavrenko, Matt Schmill, Dawn Lawrie, Paul Ogilvie, David Jensen, and James Allan, 2000. Mining of Concurrent Text and Time Series. In Proceedings of the Text Mining Workshop of the Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Boston, MA, USA. [Lesk 86] Lesk, Michael. 1986. Automatic sense disambiguation: How to tell a pine cone from an ice cream cone. In Proceedings of the 1986 SIGDOC Conference, pages 24-26, New York. Association of Computing Machinery. [Leuski 01] Leuski, A. "Evaluating Document Clustering for Interactive Information Retrieval," In the Proceedings of Tenth International Conference on Information and Knowledge Management (CIKM'01), 2001. pp. 33-40. [Lewis 92] Lewis D D, 1992. Representation and Learning in Information Retrieval. Ph.D. dissertation, University of Massachusetts. [Maarek & Shaul 97] Y. S. Maarek and I. Shaul. WebCutter: A system for dynamic and tailorable site mapping. In Proceedings of the Sixth International World-Wide Web Conference, pages 713--722, Santa Clara, Ca., Apr. 1997. 6 [Maa et al. 99] Maa López, M.J., de Buenaga Rodríguez, M, Gómez Hidalgo, J.M. Using and Evaluating User Directed Summaries to Improve Information Access . Third European Conference on Research and Advanced Technology for Digital Libraries (ECDL'99), Paris, France, September 22-24, 1999. In S. Abiteboul and A.M. Vercoustre (eds.), Research and Advanced Technology for Digital Libraries. Springer-Verlag, 1999 (Lecture Notes in Computer Science, Vol. 1696), 198-214. ISBN 3-540-66558-7. [Manning & Schütze 99] Christopher D. Manning and Hinrich Schütze. 1999. Foundations of Statistical Natural Language Processing. Cambridge, MA: MIT Press. [Mrquez et al. 00] L. Mrquez, L. Padró & H. Rodríguez. A Machine Learning Approach to POS Tagging. Machine Learning, 39:1 (59-91). Kluwer Academic Publishers. April 2000. [McCallum et al. 00] Andrew McCallum, Kamal Nigam, Jason Rennie, Kristie Seymore. Automating the Contruction of Internet Portals with Machine Learning. Information Retrieval Journal, volume 3, pages 127-163. Kluwer. 2000. [Mladenic & Grobelnik 98] Mladenic, D., Grobelnik, M. (1998) Feature selection for clasification based on text hierarchy. Working notes of Learning from Text and the Web, Conference on Automated Learning and Discovery CONALD-98. [Mladenic 99] Mladenic, D., 1999. Text-learning and related intelligent agents. IEEE EXPERT, Special Issue on Applications of Intelligent Information Retrieval), JulyAugust 1999. [Mladenic98] Mladenic, D. (1998) Turning Yahoo into an Automatic Web-Page Classifier (uncompressed) Proceedings of the 13th European Conference on Aritficial Intelligence ECAI'98 (pp. 473-474). [Mooney 96] Raymond J. Mooney. Comparative Experiments on Disambiguating Word Senses: An Illustration of the Role of Bias in Machine Learning. Proceedings of the 1996 Conference on Empirical Methods in Natural Language Processing, pp. 82-91, Philadelphia, PA, May 1996. [Moore 00] Alvin Moore and Brian H. Murray, 2000. Sizing the Internet: A Cyveillance Study. [Nahm & Mooney 02] Un Yong Nahm and Raymond J. Mooney. Text Mining with Information Extraction. Proceedings of the AAAI 2002 Spring Symposium on Mining Answers from Texts and Knowledge Bases, Stanford, CA, March 2002. [Nasukawa & Nagano 01] T. Nasukawa and T. Nagano, 2001. Text analysis and knowledge mining system. IBM Systems Journal, Vo. 40, No 4. [Nigam et al. 00] Kamal Nigam, Andrew McCallum, Sebastian Thrun and Tom Mitchell. Text Classification from Labeled and Unlabeled Documents using EM. Machine Learning, 39(2/3). pp. 103-134. 2000. [Oard & Marchionini 96] Douglas W. Oard and Gary Marchionini. A conceptual framework for text filtering. Technical Report CS-TR3643, University of Maryland, College Park, MD, May 1996. [Oracle 97] Oracle, 1997. Managing Text with Oracle8TM ConText Cartridge. An Oracle Technical White Paper. 7 [Pedersen 02] Ted Pedersen. Evaluating the Effectiveness of Ensembles of Decision Trees in Disambiguating Senseval Lexical Samples. Proceedings of the Workshop on Word Sense Disambiguation: Recent Successes and Future Directions. July 11, 2002, Philadelphia. [Provost & Fawcett 01] Foster Provost and Tom Fawcett. Robust classification for imprecise environments. Machine Learning Journal, 42(3):203{231, March 2001. [Provost & Fawcett 97] Foster Provost and Tom Fawcett. Analysis and visualization of classifier performance: Comparison under imprecise class and cost distributions. In Proceedings of the Third International Conference on Knowledge Discovery and Data Mining, 1997. [Rajman & Besaçon 98] M. Rajman and Besançon R., 1998. Text Mining - Knowledge extraction from unstructured textual data. In Proc. of 6th Conference of International Federation of Classification Societies (IFCS-98), Roma (Italy), pp. 473-480. [Riloff & Lehnert 94] Riloff, E. and Lehnert, W. (1994) "Information Extraction as a Basis for High-Precision Text Classification". ACM Transactions on Information Systems , July 1994, Vol. 12, No. 3, pp. 296-333 [Riloff 95] Riloff,E. 1995. Little words can make a big difference for text classiFDWLRQ In Proceedings of SIGIR-95,18th ACM International Conference on Research and Development in Information Retrieval (Seattle, US, 1995), pp.130 ­136. [Rüger & Gauch 00] Rüger, S.M. and S E Gauch: Feature Reduction for Document Clustering and Classification. DTR 2000/8, Department of Computing, Imperial College London, September 2000. [Sahami 98] Sahami, M. 1998. Using Machine Learning to Improve Information Access. PhD Thesis, Stanford University, Computer Science Department. STAN-CS- TR-98-1615. [Sahami et al. 98] Mehran Sahami, Susan Dumais, David Heckerman, and Eric Horvitz. A bayesian approach to filtering junk e-mail. In Learning for Text Categorization: Papers from the 1998 Workshop, Madison, Wisconsin, 1998. AAAI Technical Report WS-98-05. [Salton & McGill 83] G. Salton and M. J. McGill. Introduction to Modern Information Retrieval, McGraw Hill Com- puter Science Series, New York, 1983. [Salton 89] Salton G. (1989). Automatic Text Processing: The Transformation, Analysis and Retrieval of Information by Computer. Addison-Wesley. [Schapire & Singer 00] Robert E. Schapire and Yoram Singer. BoosTexter: A boostingbased system for text categorization. Machine Learning, 39(2/3):135-168, 2000. [Schmid 94] Helmut Schmid. Probabilistic Part-of-Speech Tagging Using Decision Trees. Proceedings of International Conference on New Methods in Language Processing. September 1994. [Scott & Matwin 99] Scott S. and Matwin S. (1999). Feature engineering for text classication. In Bratko I. and Dzeroski S. editors, Proceedings of ICML-99, 16th International Conference on Machine Learning, pages 379-388, Bled, SL. Morgan Kaufmann Publishers, San Francisco, US. [Sebastiani 02] Fabrizio Sebastiani. Machine learning in automated text categorization. ACM Computing Surveys, 2002. 8 [Semio 02] Semio Corporation, 2002. Text Mining and the Knowledge Management Space. A Semio Corporation white paper, available at http://www.dmreview.com/portal_ros.cfm?NavID=92&WhitePaperID=80&PortalID=1 7 [Steinbach et al. 00] Michael Steinbach, George Karypis and Vipin Kumar. A Comparison of Document Clustering Techniques. Text Mining Workshop, KDD, 2000 [Stevenson & Wilks 99] M. Stevenson, Y. Wilks. Combining Weak Knowledge Sources for Sense Disambiguation. Proceedings of the International Joint Conference for Artificial Intelligence (IJCAI-99). Stockholm. (1999) [Tan & Teo 98] Ah-Hwee Tan and Christine Teo. Learning User Profiles for Personalized Information Dissemination. In Proceedings, International Joint Conference on Neural Networks (IJCNN'98), Alaska, May 4-9, 1998, p183-188 [Tan 99] Ah-Hwee Tan. Text Mining: The state of the art and the challenges. In Proceddings, PAKDD'99 workshop on Knowledge Disocovery from Advanced Databases, Beijing, April 1999, p65-70. [Tjong & Buchholz 00] Erik F. Tjong Kim Sang and Sabine Buchholz, Introduction to the CoNLL-2000 Shared Task: Chunking. In: Proceedings of CoNLL-2000 and LLL- 2000, Lisbon, Portugal, 2000. [Tkach 98] Daniel Tkach. Text Mining Technology: Turning Information Into Knowledge. A White Paper from IBM, IBM Software Solutions, February 17, 1998 [Tzeras & Hartmann 93] Tzeras,K.and Hartmann,S.1993.Automatic indexing based on Bayesian inference networks. In Proceedings of SIGIR-93,16th ACM International Conference on Research and Development in Information Retrieval (Pittsburgh, US, 1993), pp.22 ­34. [Urea et al. 01] Urea López, L.A., de Buenaga Rodríguez, M., Gómez Hidalgo, J.M. Integrating linguistic resources in TC through WSD , Computers and the Humanities, Volume 35, Issue 2, May 2001. [Vaithyanathan & Dom 99] Vaithyanathan, S. and Byron Dom. "Model Selection in Unsupervised Learning With Applications To Document Clustering". The Sixteenth International Conference on Machine Learning (ICML-99). June 27-30, 1999 in Bled, Slovenia. Proceedings: Edited by I. Brakto and S. Dzeroski; Published by Morgan Kaufman [Vergne 00] Jacques Vergne. Trends in Robust Parsing. A tutorial presented in Coling 2000. [Vossen 98] Vossen, P. (ed) 1998 EuroWordNet: A Multilingual Database with Lexical Semantic Networks, Kluwer Academic Publishers, Dordrecht [Wise et al. 95] J. A. Wise, J. Thomas, K. Pennock, D. Lantrip, M. Pottier, A. Schur, and V. Crow. Visualizing the Non-Visual: Spatial Analysis and Interaction with Information from Text Documents. In N. Gershon and S. Eick, editors, IEEE Information Visualization '95. IEEE, Oct. 1995. ISBN 0-8186-7201-3. [Witten & Frank 99] Witten I. H. and Frank E. (1999). Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations. Morgan Kaufmann. 9 [Wu et al. 01] Wu, M., Michael Fuller, and Ross Wilkinson. Using Clustering and Classification Approaches in Interactive Retrieval. In Information Processing & Management, pp. 459-484, 37(3), 2001 [Yang & Liu 99] Yiming Yang and Xin Liu A re-examination of text categorization methods. Proceedings of ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR'99, pp 42--49), 1999. [Yang & Pedersen 97] Yang, Y., Pedersen J.P. A Comparative Study on Feature Selection in Text Categorization Proceedings of the Fourteenth International Conference on Machine Learning (ICML'97), 1997, pp412-420. [Yang 99] Yiming Yang An evaluation of statistical approaches to text categorization. Journal of Information Retrieval, Vol 1, No. 1/2, pp 67--88, 1999. [Yang et al. 99] Yiming Yang, Jaime Carbonell, Ralf Brown, Thomas Pierce, Brian T. Archibald, Xin Liu Learning Approaches for Detecting and Tracking News Events. IEEE Intelligent Systems: Special Issue on Applications of Intelligent Information Retrieval,Vol. 14(4), pp32-43, July/August 1999. [Yarowsky 92] Yarowsky, D. `` Word-Sense Disambiguation Using Statistical Models of Roget's Categories Trained on Large Corpora.''In Proceedings, COLING-92. Nantes, pp. 454-460, 1992. [Zhao & Karypis 00] Ying Zhao and George Karypis. Criterion Functions for Document Clustering: Experiments and Analysis. TR# 01-40, Department of Computer Science & Engineering, University of Minnesota, Minneapolis, 2000.