Artificial Intelligence: Web Applications

1
Artificial Intelligence Web Applications

• Lecture 16

• Knowledge and the Web
• Semantic Web
• Information Extraction

2
Knowledge and the WWW

• The WWW as it stands is primarily just a
  collection of documents with links between them.
• It is up to the human reader to determine meaning
  and knowledge from the documents.
• This puts a lot of burden on the human reader -
  How can they best find the knowledge that they
  require from the huge volume of diverse texts?

3
Searching for Information

• Currently search engines are used.
• The users query is matched against possible
  documents - documents containing query terms are
  returned.
• Based on huge index containing words/sequences
  that might occur in the text part of a document.
• Largely relies on literal matches of text
  fragments.

4
Knowledge, Information and Text

• But the same knowledge may expressed through many
  different words..
• Matching on words is not really adequate. Really
  want to match on meaning.
• Also leaves much work for user - skim through all
  the documents returned to find one which has
  knowledge required.

5
Towards more meaningful search

• This has motivated work in two areas
• The Semantic Web aims to provide
  representations and linkage of documents at the
  level of meaning, allowing more knowledge-based
  traversal and search.
• Work in Natural Language has shifted to the
  problem of extracting information from many
  texts, and creating summaries and tables.

6
Semantic Web

• Semantic Web based on two main things
• Metadata provides knowledge ABOUT documents
  (e.g., author, topic, date).
• Ontologies provide information about the
  concepts that may occur in documents, and how
  they are related.

7
Metadata

• Documents should have associated with them
  structured representation about the document.
• Current standard is RDF Resource Description
  Framework.
• An RDF Record is network based - analogous to
  semantic networks.

HW
worksFor
author
topic
Doc1
AI
Fred
Phone
1224
8
Metadata Schemas

• Need some control so that everyone writes
  metadata specifications to same standard.
• Metadata schemas have this role. They specify
• What the allowed values are for different
  relations (e.g., phone number must take a
  number).
• What the fixed relations between different
  concepts are (e.g., AI is a subtopic of CS).

9
Metadata and Ontologies

• Helpful if all metadata descriptions are based on
  the same terms/concepts, and agreed relations
  between these concepts.
• An agreed description of concepts in a domain and
  their relations is called an ontology.
• E.g., ontology could include fact that CS has sub
  topics AI, DB, Graphics
• This is only one conceptualisation. Could
  divide up CS into different topics (e.g, software
  vs hardware).
• Using agreed conceptualisation helps in finding
  meaningful relations between documents.

10
Semantic Web

• Using Metadata and Ontologies, we have a
  representation of documents at the level of
  concepts, rather than strings.
• We can find precise relations between the
  concepts in two documents (e.g., the authors work
  at the same place), rather than just a
  meaningless link.

11
Web Services

• Rich metadata also forms the basis for developing
  web services.
• WSDL (web service definition language) provides
  spec. of service service profile and process
  model
• functionality, use, cost, quality, location etc.
• inputs/outputs, invocation mechanisms, platform.
• Ontology languages (DAML-S, OWL-S) provide
  definitions/semantics of terms used in service
  descriptions
• Services can be automatically found and composed
  to meet user needs and run over Web.
• AI planning sometimes used for this.

12
Information Extraction

• Metadata has to be written by someone, so load on
  document authors.
• It also doesnt contain much of the actual
  knowledge in the document, just statements
  about the document.
• So interest in using NL techniques to skim
  documents and extract key facts.

13
Information Extraction

• Information Extraction programs analyse documents
  on a particular subject, and extract specified
  data. E.g., goals scored and teams from football
  report
• Use natural language understanding methods
  augmented with rule-based techniques.
• Dont aim to get a full semantic representation
  of every statement in document, just parse it
  enough to be able to apply rough rules.
• E.g., to find teams, find a noun phrase followed
  by a verb phrase mentioning play.

14
Knowbots and Infobots

• Idea of intelligent agents also applied to Web.
• Knowbots/infobots are agents whose job it is to
  find information/knowledge on behalf of the user.
• They know about the users interests and
  information needs, and traverse the Web acting on
  behalf of the user.
• They may have powers to negotiate for the user,
  requesting other agents to act.

15
AI and Web

• So.. Current Web is about documents with text and
  other media, and single kind of link between
  documents.
• Vision is web of knowledge, with links based on
  relations between concepts.
• Need to think at level of concepts, not just text
  strings.
• Ideas from knowledge representation contribute to
  this conceptual level representation.
• NL contributes to extracting information from
  text.
• Intelligent agents act on behalf of user to
  gather relevant knowledge.

16
Summary

• AI now applied to managing information on the
  Web.
• We need more than just linked texts, but
  authenticated Web of knowledge.
• We need to extract knowledge from text, and
  represent more meaningful relations between
  documents.
• A mixture of technologies contributes to this..
• But still at relatively early stages.