Ontology-based Web Information Extraction in Practice

Transcription

1 Ontology-based Web Information Extraction in Practice erecruitment etourism - eprocurement Japan-Austria Joint Workshop on ICT Tokyo, October 18-19, 2010 Institute for Application Oriented Knowledge Processing a.univ.-prof. Dr. DI Birgit Pröll bproell@faw.jku.at

2 Contents Motivation Web Information Extraction (WebIE) by Examples General Architecture Web Crawler Ontology Aware WebIE Structure Analysis: Page Segementation, Table Extraction Evaluation & Manual Correction of Results Lessons Learned & Future Work Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 2

3 Web Information Extraction (WebIE) extracting structured data from Web pages accomodation s name Alpenrose phone address pool facility templates ++43 (0) A-6212 Maurach - Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 3

4 WebIE Projects in cooperation with Austrian Industry Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 4

5 Projects Requirements and Approach Taken Some WebIE pecularities in the given projects Heterogeneously designed Web pages Mixture of (semi-)structured data and full text Significant structural aspects, e.g., location of information on Web page information hidden in Web tables Information scattered over several Web pages Web site evolution WebIE Approaches Screen scraping approaches (wrapper generation) Automatically trainable systems (machine learning) Knowledge-engineering Knowledge-engineering approach approach + Web crawler + structural analysis + [Appelt et al., 1999] Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 5

6 Contents Motivation Web Information Extraction (WebIE) by Examples General Architecture Web Crawler Ontology Aware WebIE Structure Analysis: Page Segementation, Table Extraction Evaluation & Manual Correction of Results Lessons Learned & Future Work Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 6

7 Overall Architecture Pre-Processing Information Extraction Post-Processing Crawler IE-Pipeline (GATE *) Output Web sites Tokenizer Sentence- Splitter (e.g.) Gazetteer- Lists Ontology- Plugin Transducers Annotated Web pages <?xml version=1.0 > <masterdata> <accname> Alpengasthof XYZ </accname> <adress> Baumbachstr Linz </adress>.. </masterdata> XML Gazetteer lists Rules Knowledge Base Domain Ontology *) [Cunningham et al, 2006] Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 7

8 Web Crawler Collects relevant Web pages Classifies Web pages Home page, price pages, location pages, etc. Based on Support Vector Machine Recognises language Using meta-tags and an n-gram based algorithm Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 8

9 Overall Architecture Pre-Processing Types of annotations - syntactical, morphological - ontological Information - structural Extraction - relevance judging Post-Processing Crawler IE-Pipeline (GATE) Output Web sites Tokenizer Sentence- Splitter (e.g.) Gazetteer- Lists Ontology- Plugin Transducers Annotated Web pages <?xml version=1.0 > <masterdata> <accname> Alpengasthof XYZ </accname> <adress> Baumbachstr Linz </adress>.. </masterdata> XML Gazetteer lists Rules Knowledge Base Domain Ontology Manual Evaluation Manual Correction Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 9

10 Regular Expressions & Gazetteer Lookup Rule: Phone1 ( {Token.string=="+"} {Token.kind==number} ({SpaceToken.kind==space})* {Token.string=="("} {Token.kind==number} {Token.string==")"} (({SpaceToken.kind==space})* {Token.kind==number})+ ):phone --> :phone.myphone={} Gazetteer list phone keywords Phone Telephone Tel. Tel: Tel.: Telefon Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 10

11 Ontology-Aware Entity Recognition (1/2) hotel amenities pool facilities rdfs:subclassof swimming pool rdfs:label lang=en owl:synonym Schwimmbad Hallenbad rdfs:label lang=de Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 11

12 Ontology-Aware Entity Recognition (2/2) hotel amenities rdfs:subclassof pool facilities pool attributes We offer a wonderful 2500m2 wellness area, lead by a trained wellness team. Indoor swimming pools, new heated natural outdoor pool with sandy beach, open air whirlpool with a wonderful view of lake Caldaro, large sauna world, and our private beach directly at lake Caldaro, full fill all wishes! pool hasattibute (ObjectProperty) heated Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 12

13 Structure Analysis: Web Page Segmentation Top part templates job title Senior Java Developer IT skills + level JAVA + perfect MySQL + basic Content part operation area SW programming, testing language skills English fluently contact - powered by Typo3 Bottom part [Debnath et al., 2005] [Chakrabarti et al., 2007] Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 13

14 Structure Analysis: Block Identification Content part Block Responsibilities Block Requirements Block Offer Block Block Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 14

15 Structure Analysis: Table Data Extraction in Marlies [Yang et al., 2002] [Gatterbauer et al., 2007] Japan-Austria Japanese-Austrian Joint Workshop on Natural ICT, Oct. Language 18-19, 2010 and Spatio-Temporal Birgit Pröll Information, 1st-2nd Oct Birgit Pröll 15

16 Structure Analysis: Table Data Extraction in TourIE Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 16

17 Contents Motivation Web Information Extraction (WebIE) by Examples General Architecture Web Crawler Ontology Aware WebIE Structure Analysis: Page Segementation, Table Extraction Evaluation & Manual Correction of Results Lessons Learned & Future Work Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 17

18 Evaluation: TourIE Evaluation results were satisfactory with respect to the preliminary study. Pool facility extraction quality was poor because of incomplete ontology. Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 18

19 Evaluation: JobOlize 1,2000 1,0000 0,8000 0,6000 0,4000 0,2000 0,0000 Precision Recall F-Meas. Precision Recall F-Meas. Precision Recall F-Meas. Operation Area IT-Skill Language Skill Context-driven Extraction Non Context-driven Extraction Page segmentation & block identification considerably rises precision. Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 19

20 Evaluation: Marlies Marlies Ontology Classes: 2313 Instances: 2661 Assignments of object properties to instances: Gesamt Preliminary results for recall: 1,00 0,90 0,80 0,70 0,60 % 0,50 0,40 0,30 0,20 0,10 0,00 Gesamt Name Adresse Mail Telefon Fax UID Attribute Bezeichnung Material Hauptmaß H-Einheit H-Wert XYZ Material Work in progress (e.g., table extraction). Templates Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 20

21 Manual Correction via Rich Client GUI Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 21

22 Contents Motivation Web Information Extraction (WebIE) by Examples General Architecture Web Crawler Ontology Aware WebIE Structure Analysis: Page Segementation, Table Extraction Evaluation & Manual Correction of Results Lessons Learned & Future Work Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 22

23 Lessons Learned Today s Web pages do not adhere to standards or semantic Web proposals. Only a few RDF resouces available; proposed microformats rarely used Poor HTML, e.g, tables used for layout purposes Web 2.0 coded Web pages in progress; content-based image retrieval & OCR Development & maintenance of knowledge-based WebIE systems is expensive. Domain experts & knowledge engineers are needed. Rule-coding is tedious and errorprone. Evaluation of numerous methods & algorithms; multiplied due to multilnguality Manual evaluation is time consuming. WebIE performance considerably depends on quality of domain ontology. We have to observe (evolving) legal issues Robots exclusion standard, Sitemap etc. Further processing of extracted data Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 23

24 Future Work: Ontosophia Ontology-driven IE Supported by (Semi-) Automatic Corrective Feedback (1) Extraction Domain Ontology Documents Ontology Optimization Ontology Learning, -Population, -Evaluation (2) Ontology Optimization Ontology-driven Information Extraction Extraction Domain Ontology Information Extraction Pipeline Rule-Generator IE-Evaluation (4) Evaluation Support Ontology Lookup Document Annotation Template Filling Domain- Expert(s) IE-Rules Ontology Lookup Visualization of IE-Process Error Trace Back (3) Visual Error Trace Back Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 24 Validation GUI

25 Thank you for your Attention! Acknowledgements to: Christina Stefan Christina Michael Feilmayr Parzer Buttinger Guttenbrunner Japan-Austria Joint Workshop on ICT, Oct , 2010 Birgit Pröll 25