[[page created automatically from word-processed document; for original see: Postscript version]]

(continuation slide:)

From search engine to hypertext:

a spectrum of possibilities

Peter Brown

Computer Science

Univ. of Exeter, UK

Background to work

• Links with Southampton University.
• These are early ideas.

Extremes of retrieval

• Unconstrained: search engine
• Constrained: a hypertext page including, say, 5 links to other pages that supply further information.
• We often want retrieval between these extremes ..
• .. especially when using a small screen.

Points on the spectrum

• 1: Full IR, e.g. search engine.
• 2: Information Filtering, which is like IR but:
  • user's query is supplied in advance.
  • delivery of documents is proactive.
• 3: Site search on the web; more generally searching a sub-hierarchy.

Points on the spectrum (cont.)

• 4: Context-aware retrieval:
  • may be proactive, e.g. delivery of information as tourist nears a site.
  • may be interactive, but query is wholely or partly supplied automatically.
• 5: User-tailored retrieval:
  • similar in principle to previous, but with tailoring of information to a user (e.g. preferences, past history, bookmarks).
• 6: Dynamic hypertext links:
  • source and/or destination of links are not fixed, but are calculated by a function.

Points on the spectrum (cont.)

• 7: Composite retrieval, e.g.:

  (1)

  do an ordinary retrieval.

  (2)

  apply one or more `filters' to the results of (1).

  Perhaps carry this further and have several autonomous retrieval agents + a consolidator.

• 8: simple web links:
  • the retrieval has been done in advance by the author, who has provided links to the most relevant documents.
  • author has a piece of context: the web page the reader is looking at.

Underlying mechanism: a three-stage retrieval process

(1)

specification of (1a) document collection and (1b) query.

(2)

issue of query to retrieval engine; delivery of relevant documents.

(3)

selection by user of a retrieved document.

... plus possible further step of marking relevant document fragments, and explaining why each is relevant.

• There may be iterations round these stages, e.g. building and using a cache involves iterating over (1) and (2) above.

Timing

• For large amounts of information, some elements of (1) to (3) above must be done in advance in order to get acceptable performance.
• .. but subsequent change then causes problems.

Dimensions of the spectrum

• Queries: permitted forms, who specifies, who issues.
• Document collection: who specifies.
• Iterations.
• Timing: which stages done in advance.
• Delivery: proactive/interactive.

• Most dimensions involve a choice between freedom and constraint.
• Current applications typically pick certain points on the spectra, but provide some mechanisms to ease/impose constraints.

Example: easing of constraints within hypertext system

• Use of associated glossary or dictionary.
• `Find me a web page like this one'.
• Just-in-time retrieval agents: annotating the current document with automatically-created links to relevant material; also Autonomy products.
• More generally, using intensional rather than extensional links.

Proposal: the read/write interface

• Do not distinguish between browsers (for reading) and editors/word-processors (for writing). Thus you have a browser that allows annotation, deletion, highlighting, etc., by the user.
• Current read/write document is called the document of interest, and is one part of the user's context.
• The document of interest can be used as a basis for retrieving further information the user wants.

Proposed future application: SUPERIF

• Information Filtering, but documents are found by a resource-discovery agent, which knows about user interests.
• Uses read/write interface.
• Queries are adapted automatically according to how the user reacts to the documents delivered; thus queries evolve slowly over time.
• Caters for different user `hats'.
• Aims to provide proactive delivery of all the new information the user needs.

Proposed future application: SUPERIR

• SUPERIR is an IR system that only does context-aware retrieval (but document of interest is part of this context).
• Uses read/write interface.
• Has a set of predefined document collections prepared in advance (like set of web pages used by search engines).

Difference between SUPERIF and SUPERIR

• Query known in advance vs. document collection known in advance.

Conclusions

• We have tried to find automatic processes that help find relevant documents without user intervention.
• One key to this is the use of context.
• Another key is the read/write interface.
• Some operations must be done in advance.
• The more automation there is, the more a user needs a facility to find out why a particular document has been delivered.