1. PAUL ALEXANDRU CHIRITA STEFANIA COSTACHE SIEGFRIED HANDSCHUH WOLFGANG NEJDL 1* L3S RESEARCH CENTER 2* NATIONAL UNIVERSITY OF IRELAND PROCEEDINGS OF THE 16 TH INTERNATIONAL CONFERENCE ON WORLD WIDE WEB, 2007 SESSION: SEMANTIC WEB AND WEB 2.0 PTAG: Large Scale Automatic Generation of Personalized Annotation TAGs for the Web 1

2. Outline Abstract Introduction Previous Work Automatic Personalized Web Annotations Experimental Results Conclusions Future Work Comments 2

3. Abstract The success of the Semantic Web depends on the availability of Web pages annotated with metadata In this paper they propose P-TAG, a method which automatically generates personalized tags for Web pages  produces keywords relevant to its textual content  also to the data residing on the surfer’s Desktop Empirical evaluations with several algorithms pursuing this approach showed very promising results 3

4. Introduction (1/3) The Semantic Web a vision of a future Web of machine- understandable documents and data Annotations are the main instrument, which enrich content with metadata in order to ease its automatic processing  The problem of traditional manual or semi-automatic annotation  Alternative method: tagging 4

5. Introduction (2/3) Why automatic tagging?  Webpage are growth very fast  Recommendation Why personalization?  Automatically generated tags have the drawback of presenting only a generic view 5

6. Introduction (3/3) Problems of user profile  These profiles are laborious to create and need constant maintenance in order to reflect the changing interest of the user Personal Desktop usually contains a very rich document corpus of personal information  Can and should be exploited for user personalization 6

7. Previous work (1/2) 7 - Generating annotations for web Brooks and Montanez [4]  analyzed the effectiveness of tags for classifying blog entries  and found that manual tags are less effective content descriptors than automated ones Cimiano et.al. [10, 11]  Proposed PANKOW (Pattern-based Annotation through Knowledge on the Web)  Employs an unsupervised, pattern-oriented approach to categorize an instance with respect to a given ontology  C-PANKOW: enhanced version of PANKOW  It requires an input ontology and output instances of the ontological concepts  Annotation is always directly rooted on the text of the web page

8. Previous work (2/2) 8 - Generating annotations for web (cont’d) Dill et. al. [14]  Present a platform for large-scale text analytics and automatic semantic tagging  The system spots knows terms in a webpage and relates it to existing instances of a given ontology - Text Mining for Keywords Extraction - Text Mining for Keywords Association

9. Automatic personalized web annotations (1/4) 9 Three approaches to generate personalized web page annotations  Document Oriented Extraction  Keyword Oriented Extraction  Hybrid Extraction

10. Automatic personalized web annotations (2/4) 10 Document Oriented Extraction

11. Automatic personalized web annotations (3/4) 11 Keyword Oriented Extraction

12. Automatic personalized web annotations (4/4) 12 Hybrid Extraction

13. Experimental 13 Experimental Setup  Documents set of personal desktop  E-mails 、 Web cache documents 、 all files (user selected paths)  For the annotation, the input web page were categorized  Small (below 4KB)  Medium (between 4KB and 32KB)  Large (more than 32KB)  Total of 96 web pages were used as input to be annotated  Over 2000 resulted annotations  Each proposed keyword was rated 0 (not relevant) or 1 (relevant)  Measured the quality of the produced annotations using precision  The precision at level K (P@K)

14. Experimental Results (1/5) 14 Document Oriented Extraction Small web pages Medium web pages Large web pages

15. Experimental Results (2/5) 15 Keyword Oriented Extraction Small web pages Medium web pages Large web pages

16. Experimental Results (3/5) 16 Hybrid Oriented Extraction Small web pages Medium web pages Large web pages

17. Experimental Results (4/5) 17 Precision at the first three output annotations for the best methods of each category

18. Experimental Results (5/5) 18 Examples of annotations

19. Applications 19 Personalized Web Search Web Recommendations for Desktop Tasks Ontology Learning

20. Conclusions 20 Our technique overcomes the burden of manual tagging The system does not require any manual definition of interest profiles The system proposes a more diverse range of tags which are closer to the personal viewpoint of the user The results produced provide a high user satisfaction

21. Future Work 21 A shared server approach that supports social tagging  Diversity  Keywords are generated from millions of sources  Scalability  High utility for web search, analytics and advertising  Instant update

22. Comments 22 In regard to the automatic tags generation, the existing tools are good enough to implement the system Tag recommendation is a good incentive for user to give tags Automatic tagging are aids, for the social network on the web, user’s tags represented a comprehension of “what the people is”

23. Finding Similar Documents 23 Cosine Similarity  Based on TFxIDF  The weight of terms calculated from Vectors of two documents For all terms of two documents Weights of term t for two documents

24. Extracting Keywords from Documents 24 Keyword extraction algorithms usually take a text document as input and then return a list of keywords Each keyword has associated a value representing the confidence

25. Extracting Keywords from Documents 25 For keyword extraction, they use the following methods Term Frequency Lexical Compounds Sentence Selection Document Frequency

26. Term Frequency 26 This is necessary especially for longer documents, because more informative terms tend to appear towards beginning Number of terms in the document Position of the first appearance of the term

27. Lexical Compounds 27 Noun analysis is the simplest approach for lexical compound  Step1: part-of-speech tagging for the document  Step2: finding the pattern of { adjective?, noun+ }  Step3: ordering the patterns by frequency Zero or one One or more

28. Sentence Selection 28 This technique builds upon sentence oriented document summarization Ranking the document sentences according to their salience score [26] Number of significant words in the sentence Total number of words in the sentence * Significant word Position score Optional parameter Number of query terms present in a sentence Number of terms in a query

29. Sentence Selection 29 Significant word Number of sentences in the document

30. Finding of Similar Keyword 30 For find related keywords, they use the following methods Term Co-occurrence Statistics Thesaurus Based Extraction

31. Term Co-occurrence Statistics 31 Extracted keywords from web page

32. Similarity Coefficients 32 Cosine similarity Mutual Information Likelihood Ratio

33. Thesaurus Based Extraction 33