1. Methods for the Automatic Construction of Topic Maps Eric Freese, Senior Consultant ISOGEN International

2. Agenda Automatic Construction from Structured Documents Automatic Construction from Unstructured Documents

3. Contextual Harvesting Markup can provide clues about the information within a document Largely dependent on semantic markup Takes advantage of nesting within elements Rules can be developed for harvesting data to build topic map constructs –Rules could then be applied to similar types of documents

4. DTD/Schema Development to Support Harvesting DTDs like HTML are mostly useless to a harvesting system Flat structures make associations between elements more difficult New DTD/schema development should take possible knowledge harvesting into account

5. Content-Based Harvesting Combination of contextual and natural language harvesting Text is parsed and clues within the text are used to harvest knowledge. HTML documents where labels are included in the text could be processed this way

6. NLP Strategies Named Entity recognition –A list of entities (people, companies, places, etc.) is defined –Programs parse a corpus of information to identify entities –Limited to the completeness of the entity list

7. NLP Strategies – cont. Concept extraction –A list of key words can be defined much like the named entity strategy –Common strings may also be identified and suggested as new concepts –More processing intensive than named entity

8. NLP Strategies – cont. Taxonomic classification –Documents are analysed and classified according to a human-defined taxonomy –Specialized programs must be developed that are able to understand the taxonomy Must also be able to process synonyms and related concepts

9. NLP Strategies – cont. Discourse analysis –Programs are developed that attempt to understand the meaning of a text –Analyze the parts of speech using a lexicons and rules to attempt to derive the meanings and usages of words

10. Steps in Processing Natural Language Tokenization Part of Speech Tagging Bracketing Identification of useful structures

11. Tokenization GOAL: Prepare text for processing by a natural language processing system Flowing paragraph text is broken into sentence units Sentence units are broken into word tokens Word tokens are prepared for part of speech processing –Contractions and other constructs receive special processing –n’t, ’s

12. Part of Speech (POS) Tagging GOAL: Identify the part(s) of speech for each word token Lexicon – list of words with the possible POS tags for each word EXAMPLE: –sound - NNP JJ NN VB The Sound of Music, Puget Sound a sound decision the sound of silence sound the alarm

13. POS Tagging – cont. POS tagging is difficult Exception processing often required for phrases and grouped words EXAMPLE: –Time flies like an arrow. –Fruit flies like an apple.

14. Bracketing GOAL: Identify groupings of words into phrases and the hierarchical relationship of phrases to one another A set of rules is used to identify how different parts of speech and phrases can be combined to form larger phrases. EXAMPLE: –[NP, DT, JJ, NN] –Noun phrase can consist of a determiner followed by adjective followed by a noun

15. Benefits of the phased approach The separation of functions allows this approach to be applied to any language. –A lexicon is developed for the language –Rules for language construction are defined –Generic engine is able to process the data The separation of the lexicon and the rules base allows the model to be modified/improved as the corpus of text grows.

16. Putting it all together EXAMPLE: The red ball rolled down the hill. Tokenization and POS tagging –The DT –red JJ NNP –ball NN –rolled VBD VBN –down RB IN RBR VBP JJ NN RP –hill NN

17. Putting it all together – cont. EXAMPLE: The red ball rolled down the hill. Bracketing rules –[S, NP, VP] –[NP, DT, JJ, NN] –[VP, VBD, PP] –[PP, IN, NP] –[NP, DT, NN] RESULT (using XML for bracketing): The red ball rolled down the hill

18. Harvesting Considerations GIGO rule in effect The harvesting process only hastens topic map construction Only some of the topic map merging rules are applicable –Limited prospect of meaningful subject identities Humans must still participate in the process of knowledge organization in order to maintain quality –Selective inclusion in the topic map/knowledge base

19. Q & A Questions or comments welcome at: ISOGEN International 1611 W. County Road B, Suite 204 St. Paul, MN 55113 USA Voice: 1.651.636.9180 - Fax: 1.651.636.9191 eric@isogen.com www.isogen.com

20. Demonstrations SemanText – open source Ontopia Knowledge Suite – commercial

21. Harvesting Knowledge using SemanText  Contextual Harvesting Natural Language-Based Harvesting

22. Contextual Harvesting in SemanText  XML markup is used as in NLP to identify document structures Users write rules for harvesting information into topic maps structures

23. Harvesting XML Content Data can be harvested from any XML document (including RDF) into a topic map without the need to develop specialized programs Specification language is based on Xpath In future, will also record the location from which the data was harvested as an occurrence

24. Harvesting RDF Content RDF structures can be harvested in order to build topic maps Demonstrates the possibility of interoperability between the two models Rules can be established for flavors of RDF (e.g. Dublin Core, DAML/OIL) –Allows any document using the tagging scheme to be harvested Only binary associations can be generated

25. NLP in SemanText  The initial lexicon included within SemanText  is based on a lexicon derived from the Penn Treebank tagging of the Brown corpus (1 million+ words) and a very large sample from the Wall Street Journal (approx. 3 million words) SemanText  provides the ability to identify and add new words to the lexicon through a GUI.

26. NLP in SemanText  – cont. SemanText  uses a public domain parser to tokenize flowing text and identify the appropriate POS for each word token based on a set of bracketing rules. XML is used to denote the bracketing This XML markup allows SemanText  to process natural language using its contextual-based harvesting capability