Recent Trends in Text Mining Girish Keswani gkeswani@micron.com Text Mining? What? Why? Data Mining on Text Data Information Retrieval Confusion Set Disambiguation Topic Distillation How? Data Mining Organization Text Mining Algorithms Jargon Used Background Data Modeling, Text Classification, and Text Clustering Applications Experiments {NBC, NN and ssFCM} Further work References Text Mining Algorithms Classification Algorithms Naïve Bayes Classifier Decision Trees Neural Networks Clustering Algorithms EM Algorithms Fuzzy Jargon DM: Data Mining IR: Information Retrieval NBC: Naïve Bayes Classifier EM: Expectation Maximization NN: Neural Networks ssFCM: Semi-Supervised Fuzzy CMeans Labeled Data (Training Data) Unlabeled Data Test Data Background: Modeling Vector Space Model Background: Modeling Generative Models of Data [13] : Probabilistic “to generate a document, a class is first selected based on its prior probability and then a document is generated using the parameters of the chosen class distribution” NBC and EM Algorithms are based on this model Importance of Unlabeled Data? D A B G E F Labeled Data Unlabeled Data Test Data C Provides access to feature distribution in set F using joint probability distributions How to make use of Unlabeled Data? How to make use of Unlabeled Data? Experimental Results [1] Using NBC, EM and ssFCM Experimental Results [2] Using NBC and EM Extensions and Variants of these approaches Authors in [6] propose a concept of Class Distribution Constraint matrix Results on Confusion Set Disambiguation Automatic Title Generation [7]: Using EM Algorithm Non-extractive approach Relational Data [9] A collection of data with relations between entities explained is known as relational data Probabilistic Relational Models Commercial Use/Products IBM Text Analyzer [11] SAS Text Miner[12] Singular Value Decomposition Filtering Junk Email Decision Tree Based Hotmail, Yahoo Advanced Search Engines Applications: Search Engines Vivisimo Search Engine: (www.vivisimo.com) Experiments NBC Naïve Bayes Classifier Probabilistic NN Neural Networks ssFCM Semi-Supervised Fuzzy Clustering Fuzzy Datasets (20 Newsgroups Data) Sampling I: Dataset min2 min4 min6 # Features -- 9467 5685 Sampling II: Dataset Sampling Percentage Number of Features Sample25 25% 13925 Sample30 30% 15067 Sample35 35% 16737 Sample40 40% 16871 Sample45 45% 17712 Sample50 50% 19135 Vectors Sampling I Raw Data Sampling II Vectors Naïve Bayes Classifier SAMPLE Sample25 Sample30 % TRAINING % TEST ACCURACY % 20 80 34.4637 63 36 48.4945 76 23 50.9322 82 17 47.7728 86 13 48.9971 20 80 31.5436 63 36 48.0729 76 23 47.8661 82 17 50.5568 86 13 50.4587 33 66 39.1137 66 33 46.4233 77 22 48.5528 83 16 52.7383 86 13 51.2136 33 66 39.26 66 33 47.0192 77 22 48.8439 83 16 49.6907 86 13 51.6169 Naïve Bayes Classifier 55 .01 .05 .10 .25 .50 .75 .90 .95 .99 Sample30 Sample25 Accuracy % 50 45 40 35 30 Sample25 Sample30 Sample -3 -2 -1 0 1 Normal Quantile 2 3 NBC Sample25 Sample30 55 55 Accuracy % Accuracy % 50 45 40 35 50 45 40 30 20 63 76 82 86 33 66 % TRAINING 77 83 86 33 66 % TRAINING 55 55 Accuracy % Accuracy % 50 45 40 35 50 45 40 30 13 17 23 % TEST 36 80 13 16 22 % TEST ssFCM Effect of Unlabeled Data 37.5 37.5 35 35 ACCURACY % ACCURACY % Effect of Labeled Data 32.5 30 27.5 32.5 30 27.5 20 33 42 50 % LABELED 55 60 40 44 50 57 % UNLABELED 66 80 27.5 Sample sample50 sample45 sample40 sample35 sample30 sample25 ACCURACY % ssFCM 37.5 35 32.5 30 Further Work Ensemble of Classifiers [16] Further Work Knowledge Gathering from Experts E.g. 3 class Data: Input Data {C1,C2,C3} C1 Test Data ? C2 Classifier C3 References [1] “Text Classification using Semi-Supervised Fuzzy Clustering,” Girish Keswani and L.O.Hall, appeared in IEEE WCCI 2002 conference. [2] “Using Unlabeled Data to Improve Text Classification,” Kamal Paul Nigam. [3] “Text Classification from Labeled and Unlabeled Documents using EM,” Kamal Paul Nigam et al. [4] “The Value of Unlabeled Data for Classification Problems,” Tong Zhang. [5] “Learning from Partially Labeled Data,” Martin Szummer et al. [6] “Training a Naïve Bayes Classifier via the EM Algorithm with a Class Distribution Constraint,” Yoshimasa Tsuruoka and Jun’ichi Tsujii. [7] “Automatic Title Generation using EM,” Paul E. Kennedy and Alexander G. Hauptmann. [8] “Unlabeled Data can degrade Classification Performance of Generative Classifiers,” Fabio G. Cozman and Ira Cohen. [9] “Probabilistic Classification and Clustering in Relational Data,” Ben Taskar et al. [10] “Using Clustering to Boost Text Classification,” Y.C. Fang et al. [11] IBM Text Analyzer: “A decision-tree-based symbolic rule induction system for text categorization,” D.E. Johnson et al. [12] “SAS Text Miner,” Reincke [13] “Pattern Recognition,” Duda and Hart 2000 [14] “Machine Learning,” Tom Mitchell [15] “Data Mining,” Margaret Dunham [16] http://www-2.cs.cmu.edu/afs/cs/project/jair/pub/volume11/opitz99a-html/