Dialogues With Documents: Customized Interactions With Texts Information has been building for many decades Some exciting recent developments in information distribution The Internet The World Wide Web Built upon an internet foundation Brings new importance to Vannevar Bush's original vision of hypertext ("As We May Think", Atlantic Monthly, July 1945) Digital Libraries Typically built upon a WWW foundation Breathes life into H. G. Well's vision (World Brain, Doubleday, 1938)

A Pipeline for Information Traditional information flow: Augmented version showing Information Customization: Examples of information customization Bibliographic database searches Interactive data visualization (e.g. work at Pacific Northwest National Laboratory) Computer assisted foreign language learning (e.g. Berleant et al., in Computer Assisted Language Learning, 1997 no. 2) Customized access to documents A few more details:   Berleant and Berghel, "Customizing Information", Computer, Sept. & Oct. 1994

Customized Access to Documents I: HyperBrowser Occurrences of similar words are formed into hyperlinked chains Links connect "most similar sentences sharing a stem" (Except - end of a chain links to beginning) HyperBrowser: Uses preexisting Web clients as interfaces Is prescriptive (Alas, hyperlinks have just one destination) Is interactive Users follow custom pathways through the document Supports low cognitive overhead interaction Has real time navigation Has non-dialogue interaction style

Which Passage Matches the Query Best? Let M be the current best matching passage M = 1st passage in document; for each passage P in document if (P matches query better than M) M=P; Measuring degree of match - Algorithm 1 (Given query Q, passage P) Q'=getContentWords(Q); P'=getContentWords(P); return # words of Q' in P'; Measuring degree of match - Algorithm 1a . . . return # words of P' in Q'; Example: P=Dogs are dogs. Q=What are some characteristics of dogs? Algorithm 1: match value=1 Algorithm 1a: match value=2 One Solution: a more sophisticated "cosine" algorithm Problems: Many algorithm variations exist It is not always clear which is best

An N-Gram Based Strategy (Given query Q, passage P) Q'=getContentWords(Q); P'=getContentWords(P); Q''=getNgrams(Q',n); P''=getNgrams(P',n); return match(Q'', P'') (where match(Q'',P'') is e.g. a cosine measure) getNgrams(WordList, n) ngrams={ }; for each word W in WordList ngrams=ngrams + substrings(W,n); substrings("seminar",4) returns semi, imin, mina, inar We've often used n=4 Mu (1997) allowed users to set n

An N-Gram Based Strategy With Variable N (Given query Q, passage P) Q'=getContentWords(Q); P'=getContentWords(P); Q''=getNgrams(Q'); // Q''=getNgrams(Q',n); P''=getNgrams(P'); // P''=getNgrams(P',n); return match(Q'', P'') (where match(Q'',P'') is e.g. a cosine measure) getNgrams(WordList) // getNgrams(WordList, n) ngrams={ }; for each word W in WordList // ngrams=ngrams + substrings(W,n); ngrams=ngrams + substrings(W,round(0.7*length(W)));

Customized Access to Documents II: CyberBrowser (Ref.: Berghel, Berleant, Foy, & McGuire, Journal of the American Society for Information Science (JASIS), in press) List of "most important words" is automatically extracted Displays chart of which sentences contain which listed words Reader interactively, graphically specifies boolean query on listed words Displays extract of sentences conforming to query CyberBrowser: Was made into a web client (Most any document display system can be easily made into a web client!) Is nonprescriptive (to a degree!) Is interactive Has rather high cognitive overhead interface Supports real time navigation Has dialogue-like interaction

Customized Access to Documents III: MultiBrowsers Feature "multiway lookahead" Displays several subwindows Each subwindow contains a passage Each query refreshes all subwindows with new content The MultiBrowsers: Included standalone, web client, and CGI versions Are nonprescriptive Are interactive Have rather high cognitive overhead interfaces Are slow Support dialogue-like interaction Support multiway lookahead The next step: true document dialogue!

Dialogues With Documents - the Next Phase: Some Specs Should have multiway lookahead Multiple responses raise probability of a good one . . . and more than one good one is possible Fruitless queries will be less frequent Hence: each user action provides more value Should it have "split subwindows"? Retrieved passage visible while scrolling upper half Retrieved passage easy to find when scrolling lower half Variations on the basic concept possible Consider standard scrollable text boxes - Are split subwindows really better?? (They may sound better but this is unproven)

Dialogues With Documents - the Next Phase: Specs II Navigation must be easy and flexible Useful, low cognitive overhead option: - Clicking on retrieved passage makes it the next query Useful, expressive option: - Type-in box allows highly flexible queries History navigation: - Important in web browsers . . . same for document dialogue? "Sticky" subwindows retain their contents - The composition metaphor can be valuable

Dialogues With Documents - the Next Phase: Specs III Response times must be brisk An important HCI guideline - therefore: Update one or more subwindows immediately Retain each passage until its replacement is ready Use Fast CGI protocol (Or modify server via its API) Displayed content must have high quality Fill all subwindows - no blanks Avoid duplication among subwindows Have a meta-dialogue subwindow Have a clickable permuted index subwindow Include "sophisticated" retrieval algs.

Dialogues With Documents - the Next Phase: Design Strategy System closely associated with a server - so it's runnable from Netscape/Microsoft browsers System built from reusable software components - to support long term (unknowable) implementation needs System uses Fast CGI protocol - so document can be preprocessed just once per session Build index in preprocessing - user friendly and supports retrieval

Some Other Related Work I HCI and IR and . . . Automatic abstracting and extracting Examined starting in '50s Can involve passage retrieval Can be done in a custom manner Focus tends toward batch processing Focus tends toward template filling Electronic books Book length documents are a shared concern Ebooks could support document dialogue Focus tends toward hardware, currently This may change

Some Other Related Work II Meta-dialogue Examined as early as ELIZA Important in CAI Field currently unstructured True dialogue seems to require it Question answering E.g. FAQ Finder and related systems Similarities to document dialogue, but . . . Purpose is different More like web+browser than document dialogue

Dialogues With Documents - Conclusion Dialogue is more engaging that reading - because it customizes what is presented Users already scan and pick out passages - for online material (Nielsen, CACM, Jan. 1999) Works with books or sets of web documents - could handle the "too many web documents returned" issue Promises diverse, interesting, cross-cutting research activities Complements traditional reading with a new way to interact with documents