I’m at CIKM in Hong Kong this week. Today was tutorial day. I went to a tutorial with a very long title on information extraction, given by Marius Pasca from Google; as someone who doesn’t know much about the domain, I found it quite helpful and informative. The food here is excellent, even better than SIGIR in Singapore. The conference is in an exhibition centre at the airport itself. The centre is cavernous and, at least today, entirely empty apart from the few dozen CIKM tutorial attendees. Tomorrow, the marathon of the main conference starts: four parallel streams, each with three sessions a day, and each session having five papers in 110 minutes. That makes 22 minutes per paper, which is 30 seconds less than the 22.5 minutes at CIKM 2008. I’m not sure whether they’re reducing paper time by this amount every year as a matter of policy. If so, CIKM 2010 is going to be a doozy, because the shortest way to arrange 21.5 minute papers into multiple-of-five-minute sessions is to have 10-session paperspaper sessions running for three hours thirty-five minutes each.
Anyway, if anyone else is at CIKM, let me know; it would be great to catch up.
Do you have notes from Marius’s tutorial? I’d love to hear some of the details on workshop/tutorials.
Please keep blogging for those of us who can’t make it, especially the keynotes.
Pictures of the wonderful food would be particularly interesting!
This hardly sound exciting. Maybe the time has come to start asking ourselves why we show up at these conferences (other than to buffer up our c.v.s, and meet a few people we could have met in better settings).
Meanwhile, people, don’t forget to post your slides and papers online. Chances are that the people who will use your work are either not at the conference, or else, they are jet-lagged, sleepy or otherwise preoccupied. Want to have more impact? Maximize the availability of your content.
How much more effective it would be to replace these monsters of 20 papers in 2 hours (4 sessions, 5 papers each) with a series of 2-hour poster sessions. Every group of papers – and gee, the groups would no longer have all to be equal sized! – could be slotted into two such sessions, to ensure that presenters also get a chance to browse ‘clash’ posters. 40 posters, then, per two hour session; plus keynote talks; plus a few selected papers for full presentation. Lots of mingling opportunities, lots of opportunity for solid discussion of methods and results. The conference could run in a smaller venue, too.
Justin,
Switching from papers to posters is a good idea in principle, but in practice my impression is that people just aren’t that interested in reading posters. Certainly the poster session at this year’s CIKM seems poorly attended, but that may in part be because they’re running them simultaneously with the paper sessions and have them squashed into a small room.
Daniel,
Yes, this huge volume of quickly-delivered papers is not a great experience. It is almost a reduction to absurdity. I doubt that many people have the mental stamina to still be alert for the last two papers of a five-paper session. I’ll blog some more about my overall feelings about the process once the conference is over.
Jeff,
I feel a bit guilty for not having taken notes on the tutorial, but I suspect that for people with an information extraction background any notes I would have taken would have been pretty unsatisfactory. Very basically, Marius was looking at ways of identifying entities and is-a relationships between entities in free text. So you start off with your seed entities, then you find contexts in which those seed entities are used, then you find other terms or phrases that are used in the same
context, and connect those new phrases to the class of your existing ones. Well, that’s the basic area; of course, all the methods he discussed were more advanced than that.
One idea that he raised (based on his own earlier work) is that you can do this sort of extraction not just on text document collections, but on query logs, too. For instance, if your class is “Drug”, and your seed instances are “viagra, xanax, vioxx”, then you find queries like “side effects of viagra”, “side effects of xanax”, etc., from which you generalise the pattern “sides effects of X’. Then, from that, you can find other instances, e.g. “sides effects of lipitor”, and so forth. I guess a kind of obvious idea when discussed, but when Marius said in his outline that entity extraction could also be done on queries alone, alot of people (myself included) expressed surprise.
I wonder whether Marius will put his slides up on the web? They’re pretty easy to follow as a standalone resource.
William
Hi William, so what made you decide to attend CIKM?
I know don’t like going to conferences.
I like your paper by the way,
now I’m in Japan but it’s still with me.
s/I know don’t/I know you don’t/
If a poster session replaces 4 paper sessions (or 8 paper sessions), people will go to it! The poster sessions at other conferences are crowded and active. I’m not surprised attendance is suffering if people have the option of sitting in any of 4 paper sessions (not necessarily listening of course, but at least sitting down). So much conversation time lost, etc. You seem not to like the idea, but you don’t like the paper sessions either …
Tetsuya,
Hi! I had no choice: none of my co-authors were available to go to the conference!
Justin,
I did have a couple of good talks with people giving posters, it’s true. Perhaps you’re right — poster-based presentation is a better way of handling high-volume conferences than squeezing paper lengths. At this year’s CIKM, the posters were poorly run — about the only poor aspect of the conference. They were in a room that many people were unaware of, and they were so crowded together, facing each other in rows, that it was very difficult to squash in to see them. That though is an organizational issue. The larger problem is the status one. Do posters count as publications? And how many people will be willing to travel to conferences just to present a poster?