The Lair

So, I was thinking recently. Uncommon, I know. Savour the moment while it lasts. I realized in a blinding flash of the obvious, that “pass”, or more specifically “pass on” is an ambiguous construct. In fact, there are at least two completely different ideas that can be expressed with that phrase. There is a third construct, as tez points out - pass on is the formal phrase to use for kicking the bucket. But to concentrate on the Wordnet senses -

Pass on

1. To give, impart - “They pass on the parcel to their parents”
2. To relegate, defer or decline - “He passed on the unpalatable choices on offer”.

Why does this lingusitic oddity suddenly interest me? Well, it is remarkably like the famous Dinosaur comic on homographic homophonic autantonyms. Actually, even more interesting is the list of words defined as contronyms (same word, opposite meanings). via LL

Why this sudden interest in the unparseable? Because if you are trying to cajole a computer into understanding these constructs as they appear in written text, you need to figure out the context in which the words are being used. Specifically, you’d need to disambiguate the word - and, as those examples indicate, this is more difficult than you’d expect.

But humans cope with ambiguity in text all the time. Not entirely successfully, mind you. Leaving aside language for the moment, a great example for this phenomenon is found in poker. In lots of ways, poker bots vs humans represents the next leap forward for what people refer as artificial intelligence. Chess has already fallen (unless you’re a grandmaster or a very strong IM - there are already commercial chess programs that can routinely defeat human players). Draughts has fallen. Go will probably remain secure for a few more years yet, but the whole problem of computers dealing with uncertainty and incomplete information is a useful sidestep.

Poker is a great example of ambiguity in that it is impossible to be entirely certain if you’re doing the right thing when there are any number of choices. This is beautifully expressed in The Poker Chronicles -

The high fluctuations make proving any useful theory only possible in hindsight. You can mathematically examine your past results and prove that you are a winning player or a losing player to a high degree of certainty, but it takes such a large sample size that once done, it’s entirely useless. You can determine that your wins over the last year (if you played a hell of a lot of hands over that time) were outside of the range attributable to luck. You can discover your realistic minimum and maximum expectation for that time, and if the bottom of the range is above zero, you’re mathematically certain to have been a winning player.

But that was last year. The game has changed. You’ve made changes to your game, and aren’t playing the same way anymore. Your opponents are different. Maybe you moved up a limit or two so they’re a little tougher, or maybe you stayed at the same game but the field changed. You can’t prove that you are a winner, only that you were. So on a losing streak, you can’t simply turn to math or logic to console you, because it can only help explain the past. Ask it if you are playing well right now and the only answer you get is “I don’t know. Play a year and ask me again.”

And that’s the problem with trying to program something with an uncertain outcome. Identifying good heuristics for playing poker is difficult - mostly because what worked as a heuristic before could well not work anymore. A computer can already evaluate a spread better than the average player. But the rest? How things work out here could be a fertile ground for dealing with other sorts of uncertainty or ambiguity, including my personal little hobby horse - linguistics.