WWW 95 Trip report

This document is a little long, so if you want you can go to my comments on paper presentations from Tuesday, papers from Wednesday, papers from Thursday, or Alan Kay's talk to close the conference, and my summary comments.

Monday, April 10

The attendees included authors of several important Web search engines. Michael Mauldin from CMU discussed his Lycos engine, and (CS grad student) Brian Pinkerton from U of Washington talked about his WebCrawler. Martijn Koster talked about Nexor's Aliweb system. David Eichmann talked about his MORE system, which is sponsored by NASA as a byproduct of the Repository Based Software Engineering Program.

• These search services are getting quite popular. Lycos is getting so busy that it's usefulness is restricted. Mauldin mentioned that Lycos is running off a network of nine Sparc workstations, with three more on order. Yahoo is also becoming very busy, but with their new commercial backing maybe their hardware base will be able to keep pace with demand.
• The growth of the Net in general and the Web in particular means that whereas a robot could traverse the whole Web in a matter of hours 18 months ago, now it takes many days. This seems to be an inherent weakness of robot-based search indices, which in my opinion can be addressed only by running robots in parallel and pooling the results. Replication of search engines is also clearly a good idea. In a smaller, private network, then a robot that serves to keep all the servers abreast of what's new within that network might be very useful. So robot-collected metadata may be the beginnings of inter-server communication.
• Everybody seems interested in figuring ways for search engines to share information, although nobody's got the details worked out yet. One approach used by Aliweb involves having servers produce index files for their local webspace, and then sharing these index files when robots stop by.
• Robots are not themselves bad for the Net, since the ratio of volume of traffic generated by queries and the volume of traffic generated by robots is about 30:1.
• Andrzej (Andrew) Duda from IMAG talked about the Discover system. Discover provides a query routing service to over 500 WAIS servers. It found three references in WAISspace to whois++, and one apparently spurious reference to my hero Tom Swift.
• I'm looking forward to an upcoming article in IEEE Expert in which some of these search engines are compared.
• But robots aren't the end of the story! Stu Weibel from OCLC pointed out that indexing is not a viable option for truly large collections. Again, in the context of a smaller network, I'm led to wonder if a crude first pass through a corpus could be done, to be followed by construction of an ad hoc index of that subcollection for use with a set of related queries.
• This community is worried about charges of copyright infringement. If an indexing service creates a copy of a document and makes it available for you, especially if the service is robot-based and the document was found by the robot rather than submitted by its author, then maybe the service is violating the document's copyright. Lycos gets around this by abstracting the document, and indexing that, which (according to CMU's lawyers) falls under fair use.
• The query logs being collected by these search engines give an interesting picture of what people search for on the Web. The top ten query list includes Linux, and material of a sexual nature. Maulder says that his logs are invaluable as evidence of how people really search for information. That is, according to Maulder, that average users issue relatively short queries, and then browse the returned documents, as opposed to the TREC approach, in which lengthly, complicated queries are issued. In Maulder's opinion, the TREC project has spent millions of dollars trying to answer the wrong question.
• My colleague Jim Mayfield says that Maulder's data might simply be an artifact of how good the search engines on the Web are at the moment. Perhaps if the available engines could handle longer, more complex queries, people would issue them. I think there's a paradigm collision happening: at the document level, the Web still has a hypertext structure. So with the tools we have, and this hypertext structure, the user might very well adopt a short query plus a bit of browsing style of search. In a large corpus with no hypertext structure, as is the case in TREC, then document level browsing isn't nearly as attractive as a follow-up to a short query that yielded 700 possibly relevant documents.
• There were a total of about two hundred people in attendance at the eight workshops. There were about seven hundred people in attendance at the eight tutorials. Each participant was charged 150DM, or about $105. I suspect that they turned a profit.

Tuesday, April 11

The papers are listed in the order in which they appear in the printed proceedings. These proceedings are available from Elsevier. They also still seem to be available on the conference Web server.

• The paper "Towards an intelligent publishing environment", by Pitkow and Jones, discussed URLs, URNs and URCs. The "whois++" system apparently does some URN to URL mapping.
• The DEC Web toolkit paper won the conference award for Best Paper.
• My presentation of the Rowe and Nicholas paper was well-received, I think. There were maybe 200 people in the lecture hall at the time, and of course it was broadcast over the MBone, as was the rest of the conference.
• The WebMake system looks like a good start in the area of Web-based software development environments. This business of distributed software engineering has been simmering slowly for years, perhaps waiting for some effective way of tracking and resolving dependencies between different artifacts. Apparently the CCI language is used to implement client-side scripts.
• Pitkow and Recker reported on their WWW survey. Among their findings: Only 10% of Web users are female. (And I think conference attendance reflected this.)
• In the last session of the day, Quint lamented the lack of respect for SGML that HTML represents. Specifically, he objects to browsers that don't strictly enforce the HTML DTD, and to the use of HTML to describe the appearance of a document, rather than just its structure. The tone of the conference in general was tolerant of presentation information in HTML, but opposed to lone wolf extensions ala Netscape.

Wednesday, April 12

• I attended a "birds of a feather" session on Metadata and Indexing. The NCSA/OCLC workshop, which resulted in the creation of a semantic header format, was discussed at length. I personnaly am not convinced that even a detailed header will really go far towards solbing the resource discovery problem. Several white papers have been produced by an IEEE metadata group.
• The most memorable event on Wednesday was probably the presentation by Thomas Reardon of Microsoft. Reardon is the Technical Lead of the Windows 95 Networking Group, and has been deeply involved in URLs and other Web resources being treated as first-class objects in the new Windows 95 OS. Microsoft claims to be the "API company", and is content to allow others to innovate at other layers. Reardon was only mildly successful at preventing his talk from becoming too much of an advertisement for Microsoft.
• The Network Information Discovery and Retrieval panel gave people a chance to complain about the lack of metadata, and how robot-based approaches don't seem to scale well. Tim Berners-Lee knew of five different URN schemes out there, and there is now a movement towards URAs, or uniform resource agents.

Thursday, April 13

• The first paper session had several good papers. Pfeifer's paper of freeWAIS with structured fields is be quite useful. The SFGate software, which is a Web to WAIS interface, seems to be the natural client for this version of WAIS. (Xwais does not seem to be the natural client.) The Eyesight system from CNIDR is a competitor.
• There was a BOF for Sun's new Java system. Java can be described as an object-oriented language for distributed applications. HotJava is a dynamic, extensible Web browser, written in Java. By extensible, they mean importation of new Java code, called applets, from a server, to be run as part of a client. Lots of thought went into preventing applets from becoming an infobahn for viruses and trojan horses. In my opinion, if Java takes off, then Web browsers as we know them know will seem very old fashioned, and sooner rather than later. However, you need to be running Solaris on your Sparc before you can run Java. CCI and SafeTCL were mentioned as alternatives for client-side code. Since Ousterhout now works for Sun, I expect development of SafeTCL and Java to converge into a single effort.
• Pitkow gave yet another talk on Thursday. (He was a co-author on three papers at this conference.) In this paper on browsing habits of Web users, he reported that during a typical WWW session, the user visits maybe five sites, and visits perhaps seven html files per site.

Alan Kay's Closing Address

My Summary Comments

• The alphabet soup of URLs, URNs, and URCs is an active area of research.
• Nobody said a words about automated tools for converting legacy documents. Apparently the Web is growing quite fast enough with the tools we have, thank you, crude as they may be.
• Two of the keys of the Web's success: anybody can run a browser, and just about anybody can set up a simple server. Complicated servers, with multiple machines sharing one IP address, lots of proxying and caching etc., requires a lot of energy.
• Nowadays, when somebody says "object-oriented", I tend to cringe. However, Kay's argument that browsers and servers should traffic in objects, rather than (say) static HTML files, is quite compelling. If he's right, then the clients and servers of the future are going to have to exploit and implement these operations, whatever they may be. This will effect the whole architecture of the Web in ways that at the moment are not easily understood.
• We should resist efforts to standardize HTML, I think. The Web as a collaborative work environment, for training, software development, information gathering and analysis, or whatever, will accelerate as client-side and server-side scripting gets more sophisticated and robust. Reardon spoke of the Web as a platform on which other nifty applications will be developed. I wonder what those applications will be?

Last Modified: 4/19/95

by