AI Zone Admin Forum Add your forum

NEWS: Chatbots.org survey on 3000 US and UK consumers shows it is time for chatbot integration in customer service!read more..

Final 4 announced!
 
 
  [ # 16 ]

*edit* Never mind, I’ll just wait for the transcripts. I’ve got some really weird malfunctions in things too simple to be vulnerable to malfunctions. I didn’t think computers would get stage fright.

 

 
  [ # 17 ]

Yes Don there are so many possible malfunctions.
I expect some terrible “whoops” times on my part when things would go public.


@ Steve:
I hope to be able to get to Bletchley!
But I am afraid to be a bit undermined by my rather poor performance this year.

I still don’t know if I’ll be able to witness and support the event; the Live is so very soon!
Maybe I’ll be more certain in a few weeks.

You were very welcoming last year, I sincerely loved it.

 

 

 
  [ # 18 ]

With the finals coming up next saturday I don’t think we have to expect the qualifying round transcripts for some time still. So, for whom it may interest to take a peek at my AI’s performance, I’ve put the answers of my backup copy up on my blog, which should match the actual results. Along with my usual analytical commentary on the event.

Good luck to the finalists!

 

 
  [ # 19 ]

Thank you Don!


It was very interesting, well written, and makes the competition sound even more human to me.


I can’t be there this year to support the event because of money but I came here to say:
GOOD LUCK TO THE FINALISTS!

 

 
  [ # 20 ]

The transcripts of the selection questions have been posted to the AISB site.
http://www.aisb.org.uk/events/loebner-prize#Results2015

A lot of work must have gone into Cyrabot smile
My favourite transcript is Columbina for the sheer randomness of it.
Also, not sure how Robots without Borders placed 3rd from last, as it didn’t seem to respond to anything.

 

 
  [ # 21 ]

Finally! Strange how a few of my program’s answers still differ from my backup copy. I tell you, the first thing to blow up behind my back is the “Hello, my name is…” procedure. Rose seems to have had a similar issue. I wonder if maybe they typed “Hello I am Andrew” as a test to check the connection before starting for real. What do you think, Bruce?

Looks like I got the second Winograd Schema pronoun “they” = “Joe” right after all. I actually don’t know why - I mean - just as I intended, of course smile

I’m thinking maybe the transcript of Robots without Borders is incomplete rather than the bot.
Tip for Johnny: Consider answering in the negative by default when a question is about the bot and it doesn’t know. It follows the logic that if you did something, you’d probably know about it. So if you don’t know abou it, you probably didn’t do it.

 

 
  [ # 22 ]

I mentioned about the delay in posting the selection transcripts to the organisers when I was in Blectchley. They said, that they didn’t want to post them before the contest as it may prejudice the judges if they happened to see the logs. Fair enough, Mitsuku pretends to have a different name in the Loebner version and I suppose seeing her assumed name in the logs would have given it away.

 

 
  [ # 23 ]
Don Patrick - Sep 21, 2015:

Tip for Johnny: Consider answering in the negative by default when a question is about the bot and it doesn’t know. It follows the logic that if you did something, you’d probably know about it. So if you don’t know abou it, you probably didn’t do it.

Thank you for your advices. As I mentioned in another thread, I prefer to answer “I do not know” rather than giving a false answer. Unfortunately, this year the questions were very elusive, so Johnny did answer too often “I do not know,” which was not natural.

In fact, the underlying reasoning of a human is a little more complicated than that. For example, to the question “Did you see the Turing film”, it would happen something like that in the head of a human:

1) When someone sees a movie, he usually remembers.
2) I do not remember to have seen this film.
3) So I did not saw it.

Johnny is able to apply rules and deductions (maybe not exactly those ones but it could). So obviously, it is more difficult to implement than a simple “pattern” / “Template” in AIML. But I am convinced that at the end the possibilities are far more extensive. This is why I continue despite some discouraging results.

 

 
  [ # 24 ]

I understand where you’re coming from. Personally I don’t consider it “false” when it’s true: My program can only really do a handful of things and it knows about those. Anyway, Turing Tests aren’t the most suitable testing grounds for logic AI, so don’t let it discourage you.

Steve Worswick - Sep 21, 2015:

they didn’t want to post them before the contest as it may prejudice the judges if they happened to see the logs.

A wise precaution, I just wish they’d have informed us when to expect the transcripts.

I hunted down the two inexplicable answers from my transcript. Arckon’s failure to introduce himself could only have occurred if the period was not typed or received, upon which “Hi, I’m Andrew who are you?” would be considered a relative clause (e.g. “I’m the guy who programmed you?”).
And choosing “Joe” for the Winograd Schema pronoun was due a natural preference for the last-mentioned subject, but both my code and tests say it’s impossible that Arckon would have even listed singular “Joe” as a potential candidate for plural “they”. So as far as I’m concerned this didn’t happen.

 

 
  [ # 25 ]
Don Patrick - Sep 23, 2015:

Arckon’s failure to introduce himself could only have occurred if the period was not typed or received…

You should have been there on the day. Spelling and punctuation were not in abundance!

 

 
  [ # 26 ]

Indeed! I watched part of the BBC live broadcast smile and I couldn’t really judge the chatbots’ answers when I knew they could barely read the questions. I don’t know if there’s an add-on spell checker available somewhere but we should probably all install one before participating in this show.
I do apply automatic sentence splitting and question detection to compensate for absent punctuation, but sometimes a mistake is just not mine to fix. It’s an odd thing for humans to pride themselves in their ability to overlook human error, when that is also the cause of their errors.

 

 
  [ # 27 ]

So, I went and programmed a typo correction algorithm. Because it was a slow day. Turns out there are some copy-paste codes readily available for a Levenshtein Distance algoritm to compare if a misspelled word resembles a word from one’s vocabulary, perhaps of use to some here.
However, I decided to program my own specifically for accidental typos with more sophisticated probabilities based on the QWERTY layout and the fact that some kinds of typos are more probable than others. I couldn’t find much research on this after ‘76, but did find one useful fact that 80-95% of all typos differ only by 1 letter from their original (replaced, added or omitted).

Typos aside, the most interesting thing I walk away with from the qualifying round lies in the Winograd Schemas’ use of the word “if”. I had a linguistic rule that says that “A cat has a tail” is about all cats in general. Whereas “If a cat has a mouse…” is apparently only about this particular cat from this scenario, but the only syntactical difference is the word “if”. Then things got confusing. “If a bed fits in a room…” are we talking about beds in general or only this one? Or could both? “A cat had a mouse”, “A man walks into a bar”, “When a dog is faster than a cat…”, general or particular cases?

 

 
  [ # 28 ]

Don,
In the event that you have never seen it, Peter Norvig (Director of Research at Google) wrote about how to write a spell corrector.

http://norvig.com/spell-correct.html

 

 
  [ # 29 ]

Yes, I came across it and found the explanation useful, but I preferred a more lightweight solution strictly for typos. The broader approach of covering all spelling including phonetic misspellings also leaves more room for error at edge cases, to which anyone with an autocorrect function on their phone can testify.
It it interesting (in a trivial sense) to note that chatbot creators are in the best position to gather data on typos: You could monitor which letters of which words people backspace over.

 

 < 1 2
2 of 2
 
  login or register to react