Xref: utzoo comp.multimedia:641 comp.ai:9542 Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!usc!snorkelwacker.mit.edu!news.media.mit.edu!media-lab.media.mit.edu!asb From: asb@media-lab.media.mit.edu (Amy Bruckman) Newsgroups: comp.multimedia,comp.ai Subject: Re: Personalised News Systems Message-ID: <1991Jun25.201150.16394@news.media.mit.edu> Date: 25 Jun 91 20:11:50 GMT References: <13510@pt.cs.cmu.edu> Sender: news@news.media.mit.edu (USENET News System) Organization: MIT Media Laboratory Lines: 28 In article osborn@socs.uts.edu.au (Tom Osborn) writes: > >I must admit some doubts about this. I had a postgrad student working >on something a bit like this a few years ago (automatic indexing and >content addressable retrieval). Some that retrieval from headlines and >intro paragraphs is fraught with problems - the key words are there, >but so is a lot of attention attracting hype. Manual key-wording is a >possible fix, but it seems that non-experts are poor at this (they >devise keys from their own perspective well, but for readers badly). At AAAI-90 in one of the applications seminars, a company presented a rule-based system developed for Reuters to do automatic indexing of news stories coming off of the wire. They have had tremendous success with the system. It is more accurate at indexing stories than humans, because humans get bored and careless. (Evidently, indexing is tedious.) And it's much quicker: stories are indexed within an hour, if I remember correctly, instead of two days. (There is no paper about this in the proceedings. There might have been a separate set of proceedings for the applications conference; I'm not sure.) The system is very simple in its design. For example, it distinguishes the meaning of the word "lead" ("I lead them to the conference room" versus "our widgit is made of lead") by looking for the presence or absence of certain other words in the surrounding text. What is interesting about the system is not its design but its tremendous practical success. -- Amy