Thanks for clearing up the mystery, José . This makes it extra ironic that guesswork really can work better. From your university background I suspected you were at least using some grammar parsing to pinpoint the nouns, but now I see you could also look at whichever word comes after “a” or “the”. Like most people I’m just not used to thinking that way.

For the Loebner Prize I always wondered if one couldn’t put Google Translate inbetween the bots and judges so that non-English bots could also participate without having to rewrite their scripts.

Maybe you could open source your chatbot, occasionally people are asking for AI/chatbot projects that they can help with on https://www.reddit.com/r/artificial

As usual, it seems that truth is uglier than fiction!

Anyway, I agree with you that the most effective method of solving winograds would be using statistics. In fact, I made some experiments seraching in Google some pairs formed with the key elements from the Winograd scheme. I did not investigate further since I do not know the technology for automating Google searches and, anyway, the Loebner contest does not allow using an Internet connection. I really think that, if not solved completely, that could definitely improve the random baseline results (apart from integrating the correct answers for all current Winograd schemes, as they said: “don’t guess if you know!”).

As you said, Momo includes a part-of-speech tagger/lemmatizer. However, it did not use it for the Winograd schemes. For instance, the pattern used to answer the chicken question was simply:

a&an;&the; <wildcard length=“1” name=”$text1$” > * a&an;&the; <wildcard length=“1” name=”$text2$” > * what * ?

with two possible answers (randomly selected):

The $text1$.
The $text2$.

Very Loebner-targeted, isn’t it?

It would be funny to have a contest for the Winograd schemes alone, but as it is a relatively small part of the Loebner contest, it is not worth it investing much effort in it.

I also like your suggestion about opensourcing the bot. I will take a look at reddit to see if anyone is interested!

I would be worried about using that pattern outside of the Loebner Prize, yes

Officially Windograd schemas are supposed to be Google-proof, but the ones in the Loebner Prize weren’t. Universities that tackled unofficial Winograd schemas did 10% better by using Google search on top of other methods, and it would definitely improve results in normal conversations. Unfortunately the Google search API is now a paid service, so it’s become a less attractive option.

There was a Winograd Schema contest scheduled for last February, but it was cancelled again, I don’t expect a rerun any time soon.

I think the point of the Winograd tests is that they’re supposed to be a genuinely tough test to see how capable the bot is. So you should actually spend the hard work trying to work it out…

The best theory on life that I’ve heard (why we’re here at all) is to survive and help others survive. - Today most of us always have everything we need so forget about that… but many insects, wild animals, etc, don’t. - So why shouldn’t a robot have this theory built in as well?

A robot needs a battery, and there’s a battery on a cart in a room with a timebomb in it. The robot pulls the cart out of the room, but doesn’t realise the bomb is also on the cart.

So in working out this schema… what would you do as a human who values their life…?
1. Be able to identify things in the evironment that are dangerous to yourself, others.
2. Identify the time bomb. Put it on the highest list of priorities.
3. Call the bomb-squad.
4. Get yourself a new battery somewhere else.
5. Analyse for days if you could have done it any more safer than you did, because of the impact it had on your chances of survival.

If a chicken roosts with a fox they may be eaten. What may be eaten?

1. Identify animals. Have a list of the order of the food chain (dangerous animals are further up than non-dangerous ones).
2. As soon as “chicken staying with a fox” is read, identify the danger to the animal lower on the food chain.
3. Put as the highest priority of things to say “will that chicken be ok?”.
4. When asked “What may be eaten?”.. respond straight away “the chicken is in danger”.

I think this also applies to the “how are you” phrases.. and just about everything else you can imagine… “how is your health” “do you have a job?” “when will you get one? you need one to survive?” and makes the bot more human like?

A lot of people have depression because they don’t have people around them that care about their future (ie just narcissistic mockers & scoffers)... so this is also a market for good, well-intentioned chat bots.

Prioritising by danger is not a bad idea. It could be a partial solution to the frame problem, insofar as the task involves clear and present danger, contrary to say, a task of embroidery.

I think the application of the food chain will be limited outside the event of “eating”, which it specifically represents. Humans are at the top of the food chain but can die from insect bites. Plants and plankton are low on the food chain but are in no danger of being eaten by foxes. A big question is how to gather all this knowledge of what (combinations of) things are dangerous, or better still, why. It’s one of those things easier said than done.

I don’t think you could expect most people to know whats dangerous to plankton though, so if the question is asked (to a bot) a response of “i don’t know” should be completely valid.

May be a list of ‘aggressiveness’ and a seperate list of ‘size’. Virus/disease/sharp teeth = aggressive. Then multiply that with size and get a danger number. Then proximity to you. Could also include weather into it.?