Hehe! Sorry no.

Actually that gives me an idea for a team part to the competition next year….

8PLA • NET - Jun 22, 2012:

Steve, may I donate my points to Aici
so Jan may go to the next round?

That’s extremely nice of you 8Pla, but I think I’ll pass: I need a little brake. Besides, what about your bots, don’t you want them to go through?

It’s a shame Aici wasn’t working at its best, as we have had many timeouts or blank responses from it. Aici is a good bot and it should have scored higher.  Unfortunately, even if you win your last battle, you still won’t have enough points to progress to the next round. Better luck next year and thanks for entering.

Yep, Aici wasn’t in it’s best shape. More luck next time. It’s been a real good test run though.

So Jan should win the “Cursed Bot” award, I think. cheese

I think I’ll put that in the back of the closest (where it would unfortunately still be visible, empty closet and all)

The contest is moving along nicely.  Now that the scores are reaching the double digits, it’s easier to get a picture of the competitive landscape.  Like I said in my first post, the structure of Chatbot Battles keeps me coming back for more.  I’ve been checking it daily.

A few random remarks:

If it was up to me I would disallow adult-themed bots from future contests.  The contest is potentially interesting to kids, and it’s probably not the best idea to make it easy for them to find adult content. 

I’m really glad to see that some of our AIML bots are doing well, even if mine are lagging.  Good job Ron Lee and Peter Lafferty!  And congratulations to all of the botmasters percolating to the top of the league tables.

It seems that the judges have different styles.  Some of them have different conversations with the paired bots, and others do more of a side-by-side comparison by asking each bot the same questions.  From the perspective of understanding the scoring, the side-by-side comparison seems better.

Chatbot Battles has been very exciting with a number of close matches.

The final round of the group stage will pit a number of the leading contenders against each other to determine 1st/2nd place seeds going into the knockout stages. At least 3 of the groups are too close to call.

I believe this contest has been a good yardstick for the state of the art in chatbot technology. No Bot has been dominant or scored more than 15 points out of a possible 20 in the first 4 rounds.

Richard Wallace - Jun 27, 2012:

I’m really glad to see that some of our AIML bots are doing well, even if mine are lagging.

If your bots are lagging, it is only because you are in some of the toughest brackets. Luck of the draw, in a different bracket you may have taken 1st. I expect the AIML bots to be some of the toughest competitors in the knockout stages.

Talking about the style of the Judges in such competitions.
Will their “smartness” teach ChatBots to be “biased”
towards such sophisticated ways of interacting instead of us
getting chatBots that ordinary average Joe kind of people could
easily relate and like but now due to the introduced bias of the Judges
we would risk to get overly intellectual scripts that ordinary people
would feel are too high brow academic in style. Just me aimlessly curious?

Tom Anderson - Jun 30, 2012:

Will their “smartness” teach ChatBots to be “biased” towards such sophisticated ways of interacting instead of us
getting chatBots that ordinary average Joe kind of people could
easily relate

Not sure if you have seen any of the logs in the contest Tom but the majority of them are just ordinary conversations by ordinary people who are not trying to trick the bots. I instructed the judges to talk to the bots as they would a normal person instead of the Loebner type judges whose questions such as “would it hurt if I stabbed you with a towel” have gone down in AI fame.

Anyone who approaches a stranger at a party and opens a conversation with “How many legs on 3 goats?” would surely not have a conversation partner for long.

Steve Worswick - Jun 30, 2012:

Anyone who approaches a stranger at a party and opens a conversation with “How many legs on 3 goats?” would surely not have a conversation partner for long.

LMAO .. . that’s great, thanks for that Steve, that tickled my funny bone.

However, if the person was a bot developer, and knew what you were up to, the conversation could get quite protracted and involved !   . . however the probablity is .. well. . not so high.

Tom Anderson - Jun 30, 2012:

Talking about the style of the Judges in such competitions.
Will their “smartness” teach ChatBots to be “biased”?

And so goes the argument against self-learning bots.  Leaving a bot to acquire new knowledge, and to then “permanently” store that information without first reviewing it, is asking for trouble. That’s why there’s a movie rating system to protect children from being exposed to undesirable language and adult or violent scenarios.

It’s true though Victor. I was in attendance at the Loebner prize in Exeter last year and do you know the first thing one of the judges said to a bot? A hello or an introduction perhaps? Nope, he launched straight in with “how many syllables in the word balloon”.

Surely, if they are trying to decide whether a bot is human or not, they should talk to the bots as they would a fellow human?

Yeah, I suupose.  On the other hand, it is a fairly good ‘attack’.  We need to make sure our bots have a “state-of-conversation” mode.

Before anything is said could be “mode = initial”.  And when mode=initial, things like “Hi, how are you?” and the like, are ‘normal’.

But asking some form of math problem right off the bat is not normal.. and a good response would be WTF lol.. . it’s what a normal human would say.

So it amounts to tracking conversation state and also, after the input is understood, it needs to be categorized… then we check to see if that category is one of the ‘normal’ categories that should arise in the given ‘mode’.

for example, if user said “ok, now we are going to play a game ..  I will ask you a series of questions”.

then that would be mode, i don’t know, perhaps “mode=question-game”.

NOW, if we get hit by “how many syllables in the word balloon”, ok, we shouldn’t really say WTF, but give the actual value.

And if the system can’t reply, who cares.  A young child doesn’t even know what the word syllable means.  So to worry about having to have all this knowledge in the bot, is a waste of time to me.

Like seriously, are we testing the amount of knowledge in a bot, OR the ability to understand and LEARN via conversation, which should be our top priority. . not knowing tons of inconsequential stuff.

What normal person would go to a 2-year old, which has very limited knowledge, but AWESOME powers of learning and undersanding language, and ask ‘how many syllables in the word ballon’, and not getting the correct answer, write the child of as useless, or non-intelligent? only an idiot would do that.  People are missing the point here, let’s do first things first.  As Carl Sagan says ‘knowing a lot, is not the same as being smart’.  It is vastly more important to have the capacity to learn new knowledge, than simply to HAVE a lot of static knowledge that was spoon fed into the system, let’s judge the systems on the former, not the latter.

What Victor said is fascinating about bots processing grammar rules and I simply do not challenge that.
A bot that parses individual parts of speech, into noun phrases, subject, verb phrase, predicate and so forth in grammar to create its responses is a bot any robot master would drool over.

On the other hand, The Loebner Prize for A.I. is a competition, so if the questions are way too easy, like for two year old humans, then all the bots would likely finish with tie scores. A process of elimination is essential for there to be one winner.

In most robot competitions, it seems brute force rules. The bot wil the biggest hammer destroys the other bot. Well, it seems that hundreds of thousands of canned phrases is like a big hammer.  It basically amounts to brute force against a bot that may for example meticulously process grammar in the contest. Of course a bot that wins by using just grammar would be a spectacular contest to see. Luckily, The Chatbot Battles, allows up to three bots per player. So this is a great contest rule because there is very little to lose trying out new bots based on new ideas.

8PLA • NET - Jul 3, 2012:

A bot that parses individual parts of speech, into noun phrases, subject, verb phrase, predicate and so forth in grammar to create its responses is a bot any robot master would drool over.

I just recently (a few weeks ago) finished the design of the ‘conversation layer’ for my system. It parses conversation into concepts (which are in fact grounded symbols), keeps track of the conversation by keeping a queue of discussed concepts, and formulates it’s output based on the current conceptual state of the conversation and it’s short-term goals within the conversation. There are no grammar-rules involved, the system learns ‘grammar’ by statistical word-pair occurrences in learning texts (books, papers, wikipedia, etc.). So it actually learns how to formulate a response.

8PLA • NET - Jul 3, 2012:

Luckily, The Chatbot Battles, allows up to three bots per player. So this is a great contest rule because there is very little to lose trying out new bots based on new ideas.

That was the idea. I didn’t want a contest where you were forced to only enter your best work. Having 3 bots per entrant allows you to experiment a little without losing all hope.

Hans,

An abstract layer may have great potential, similar to 3D rendering just a viewport,
as opposed to the entire virtual world in a computer system… An abstract layer may
focus the power to convincingly improve a simulation in a meaningful way.

Theoretically, what may an abstract layer do separately from the structure and syntax
of words, proper names, possessives, adjectives, quantifiers, prepositional phrases,
relative clauses, phrases, noun clauses, verbs, past tense verbs, progressive verbs,
perfect verbs, perfect progressive verbs, passive verbs, adverbs, negatives, prepositional
phrases, adverbial clauses and other grammar?

Hans, can you share with us what you may have been able to elminate in an abstract layer
with no noticable difference?

8PLA • NET - Jul 4, 2012:

Hans, can you share with us what you may have been able to elminate in an abstract layer with no noticable difference?

The conceptual system uses ‘invariant representations’ of a concept. So a concept does not have any past, present or future representation. It has one invariant representation. If a concept changes based on past, present or future, it is represented as different concepts. These can be linked through the ‘knowledge representation layer’ that uses some additional constructs besides the ‘concepts’ (namely; analogies, compound concepts and concept assemblies). This also eliminates the need for predicate logic.

Communication is handled by statistical data on trigrams. Sentence formulation is handled by mapping words (atoms) to concepts (invariant representations).

That’s about as far as I can go in explaining it