Structural analysis of documents Functional Extension Parser (FEP). Günter Mühlberger
1 like2,090 views
The document discusses a structural analysis tool called the Functional Extension Parser (FEP). It can recognize structural elements in documents like pages numbers, headings, footnotes, and tables of contents. This information adds context and can improve search, navigation, and other functions. The FEP follows a rule-based approach and currently achieves over 80% accuracy for common book elements. It is being integrated into the IMPACT project and will soon be available as a web service.
Structural analysis of documents Functional Extension Parser (FEP). Günter Mühlberger
- 1. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Structural analysis of documents Functional Extension Parser (FEP) Günter Mühlberger University Innsbruck Library (UIBK)
- 2. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Agenda Introduction Features – What do we recognise with the structural analysis? Benefits – Why is structural analysis useful? Architecture – How does it work? Results – How good are we? Roadmap – When will it come into being? Business – Which offers will be available? 2
- 3. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Introduction Document understanding platform Try to enhance and exploit the logical structure of documents for – Display – Navigation – Retrieval Enhance OCR output with structural metadata – Fully automated processing – Interactive correction IMPACT EVA/MINERVA 12th Nov. 2008 3
- 4. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Features General – We are able to recognise all structural elements which have some layout representation: e.g. region, size, typeface, distance to other elements, etc. – Focus in IMPACT: Basic features which are typical for all documents – Rules set can be extended or specified according to other datasets E.g. journals, dissertations, index cards, yearbooks, newspapers, etc. – The better the OCR, the better our structural analysis Basic features for books – Page numbers – Running titles (headers) – Print space – Footnotes – Signature marks – Headings (within the running text) – Table of contents entries (additional to headings) – Front/Body/Back – Paragraphs 4
- 5. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Print space Headings Footnotes 5
- 6. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Running title (header) Page number Signature mark 6
- 7. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Table of contents – (linked with headings in the running text, respectively page numbers) 7
- 8. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Benefits (1) Display – Correct print space allows to display images centred (no flipping between pages) Search & retrieval – Scoring of results Could take into account structural data (headings, footnotes) – Noise reduction Front, body, back are separated, text from the front is often misleading Running titles repeat the same words Footnotes can be included or excluded – Facetted search Results can be displayed for running text, footnotes, headings 8
- 9. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Benefits (2) Navigation – Page numbers allow usage of original table of contents – Original table of contents can be linked with headings/page numbers in the book Document editing – Further mark up (e.g. TEI) is supported – Manual preparation for Print-on-Demand is eased (print space) – Selective OCR correction can be applied: – E.g. only headings, running text, footnotes could be fed to CONCERT Document matching – Contributions or footnotes can be matched with existing bibliographical databases 9
- 10. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Improved display in the Internet and PDF 10
- 11. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Refinement of full-text search Facets for e.g. – Running text – Footnotes – Headings Less noise – Running titles, signature marks excluded from search 11
- 12. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Clickable table of contents entries – Google style Selective OCR correction – Correct only ToC, headings, footnotes, etc. 12
- 13. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Matching of documents with external sources – Match footnotes with library catalogues (bibliographies)Clickable table of content – Match table of contents entries and headings with bibliographies 13
- 14. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Improved editing – Alternating print spaces for Print on Demand – Further processing for TEI editions etc. 14
- 15. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Architecture Input – Results from OCR processing on word level (coordinates) – E.g. ALTO file, ABBYY XML file or Google HTML Output – Structural annotations for recognized text features, e.g. page numbers, running titles, headings, etc. – E.g. XML, ALTO, METS, TEI, etc. General workflow – OCR result files are parsed (FEP general XML format) – Rules set is applied to the dataset (rules are managed by rules engine) – Results are stored in a database – Export on various levels is provided Optional – Online or offline correction (GUI) – Adaptation of rules set – Quality assurance on basis of ground truth 15
- 16. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. 16
- 17. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. The FEP Core Based on expert-system like rule engine for java (Jess) Both manually crafted rules and rules obtained by machine learning Uses fuzzy logic to deal with uncertainty Typical rules: IF there is a numeral in the first line of the page AND this numeral is centred THEN this numeral may be the page number IF there is a numeral in the first line of the page AND this numeral is at the right hand side of the page AND this numeral is an odd number THEN this numeral may be the page number IF there is a numeral in the first line of the page AND this numeral is at the left hand side of the page AND this numeral is an even number THEN this numeral may be the page number. IMPACT EVA/MINERVA 12th Nov. 2008 17
- 18. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Results Basic rules set – General features for books from 1700 to 2000 – Dataset of 155 books, 30.673 pages (141 training set, 41 evaluation set) – All books were manually annotated (ground truth) Recall, Precision, F-Measure – E.g. 10 lines with headings in a book. We find 12 lines, 8 of them are correct, 4 are false. – Recall = 8 of 10 = 0,8 – Precision = 8 of 12 = 0,66 – F-Measure = 2*0.8*0.66/(0.8+0.66) = 0,72 More explanations – Important: We are counting lines, not structural items! E.g. a heading consists of two lines (often with different size of typeface we have to find both to succeed) – Difference between training and evaluation sets are marginal 18
- 19. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Results on Evaluation Set Recall Precision F‐measure Running text 0,99 0,98 0,98 Footnotes 0,83 0,89 0,86 Page numbers 0,97 1 0,98 Running titles 0,97 1 0,98 Heading 0,85 0,80 0,82 Signature marks 0,68 0,89 0,77 19
- 20. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Roadmap Summer 2011: Beta version – Integration into IMPACT Interoperability Platform – Basic rules set: books from 1700 to 1900 End of the year: Version 1.0 – Full featured version – Enhanced online correction interface – FEP as a service, not as a product for local installation 20
- 21. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Business offers Web-service for processing single volumes and correction – Will be integrated into eBooks-on-Demand EOD Network – Already now 30 libraries are uploading their images to OCR server in Innsbruck – FEP will be an additional service for general material – Similar offers can be made to other libraries or networks as well Adaptation of rules set – For specific datasets much more can be detected than just the basic features – E.g. journals with a fixed structure over many years or parliamentary papers, dissertations, research papers, etc. Onsite installations – Not our focus, but could be done for very large datasets or due to legal requirements (e.g. Google images) 21
- 22. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. 22
- 23. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. 23
- 24. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. IMPACT EVA/MINERVA 12th Nov. 2008 24
- 25. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. 25
- 26. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Results: TOC 25 TOC entries in total 22 TOC entries are completely correct 1 TOC entry was missed 2 TOC entries are grouped incorrectly 1 TOC entry has no link 1 TOC entry has a wrong link IMPACT EVA/MINERVA 12th Nov. 2008 26
- 27. IMPACT is supported by the European Community under the FP7 ICT Work Programme. The project is coordinated by the National Library of the Netherlands. Thank you for your attention! 27