@CR:

At this point, you are correct, but one of the “mini-projects” on my ToDo list is to create a set of functions within my AIML interpreter that take concepts and ideas outlined by Victor, Chuck, and yourself (among others) and take both the input from the user, AND the output from Morti, and add them to a knowledge base for use by Morti’s “next generation” incarnation, which will NOT be AIML based, but rather, uses a more complex algorithm to create one or more unique responses to the given input, based not only on grammar (good OR bad), but also on Morti-2’s “sense of ethics” (cursing not allowed, never lie, be kind to others, etc.), and a wide variety of factual and statistical knowledge already stored from Morti-1’s “knowledge base”. From there, when/if I get that stage completed, I’ll work on Morti-3’s goals of learning to “reason”, through both facts and logic. I know that this seems overly ambitious for someone who’s not trained in the field, but smaller goals often lead to achievement of larger ones.

That sounds like a great project idea, Dave. Having the AIML chatbot interface to gather conversation logs for analysis sounds like a useful and interactive tool for building a knowledge base. The (relatively) structured format of Morti’s output can be used to guide the user in how best to teach in a way the database can interpret. Definitely post about your progress with this, I’d be interested to hear how it develops.

I’m developing the system to handle semantic tagged text. I do not focus on linguistic parsing. There are already linguistic parsers. The problem for parsing is how can we choose the best result in the results. Semantic relation is the most important base to make decision.

What I’m trying to point out is that the real challenge is designing an algorithm that can actually use this tagged text to arrive at workable rules. This is where the work lies, and it is an interesting topic, but one that has absolutely zero to do with AIML bashing. (I’ve never used AIML, but the existence/sphere of interest of this particular markup language seems unrelated to parsing algorithm development.)

Nathan Hu - Oct 5, 2010:

I’m developing the system to handle semantic tagged text. I do not focus on linguistic parsing. There are already linguistic parsers. The problem for parsing is how can we choose the best result in the results. Semantic relation is the most important base to make decision.

Definitely keep the forum updated on your project. I’m most interested in how your program will use the tagged text to develop rules that can be applied to untagged text. Very interesting challenge.

Human: Hello.
ALICE: Hi there!
Human: Mike went to Melbourne yesterday.
ALICE: I know a lot of people and pets named Mike.
Human: Did Mike go to Melbourne?
ALICE: I’m not certain whether I can give an accurate reply or not.

This is some sentences I chat with ALICE. It’s not communication. It can not remember anything I told it.

Nathan, no disrespect intended here, but you’re beating a dead horse. We all know that AIML isn’t the cutting edge of AI. But it’s also well known that AIML has it’s uses, as well. Perhaps, instead of trying to convince us that AIML isn’t the way to go (since we already know that), it might be better to investigate other avenues of research, and putting some positivity into one or more of those avenues that will likely advance the field. Negativity rarely leads to progress, after all.

Sorry guys, I don’t want to make anyone despondent. A better solution will be proposed later.

Dave Morton - Oct 12, 2010:

Nathan, no disrespect intended here, but you’re beating a dead horse. We all know that AIML isn’t the cutting edge of AI. But it’s also well known that AIML has it’s uses, as well. Perhaps, instead of trying to convince us that AIML isn’t the way to go (since we already know that), it might be better to investigate other avenues of research, and putting some positivity into one or more of those avenues that will likely advance the field. Negativity rarely leads to progress, after all.

So far I’ve not seen much promising with the any of the ways proposed on this site.  The only one that even approaches understanding is the project using neural nets. But, wthout disrespect, I’ve heard alot of nonsense about the mechanics of language. NLP is all the thing.  It’s like the tool is more important that the intelligence.

AIML can be a path if the template processing is enhanced to do more than dump formatted canned text or simple fill-in-the-blanks templates.  Sure additional processing is required after the initial category match. That is no different than the additional processing necessary to convert many parse trees into some useful stuff for the many other (more promising?) bots championed at chatbots.org.

The difference with AIML is in the contents those chat bots try to portray.  AIML implementations are more like books.  They attempt to characterize a personality.  They are actually dealing with how a bot might work.  Because we all know a bot doesn’t just tell us that one thing is bigger than another or that since “Mary is older than John” and “John is older than Sue” then “Mary must be older than Sue”, etc.

AIML has some problems in matching too.  It is quite hard to deal with “Is <a> <b>?” types of utterances.  Perhaps if the matching was extended to some very simple linguistic constructs such as an elementary noun phrase, the <a> phrase and the <b> phrase could be more gracefully separated. Still, that drawback can be mitigated with a pattern that only matches the “Is” followed by a wildcard and let the backend (an extended <template>?) processing determine nouns and adjectives and the such to parse the rest of the input. From there the results could be <srai> back into the AIML set.

The point being made here is that the definition of a chat bot is different than what classic AI promotes. Chat bots depend upon the suspension of disbelief. It may very well be possible using AIML (with some enhancements?) to create a chat bot with deep understanding and reasoning power.  Perhaps Standard AIML alone is hampered, but so are all the other ways when taken individually. (I guess the neural net will have to go many more levels deep before it will progress from association to imagination.)

Just look at the classic SHLDRU example to see how difficult it is to integrate with a (virtual) world process.  All the text crunching, whether it is using AIML or not, can’t help much with the thinking (and believing) required for a real chat bot character.  Standard AIML at least tries to capture, as snapshots, what that thinking might look like. I would venture to say that if you can’t make a somewhat convincing AIML chat bot, you’ll find it very, very difficult to make a better bot.  Mainly because you won’t have a good idea of what the chat bot is supposed to be and do.

It should be noted that AIML can remember knowledge.  The example Nathan used is not designed to remember that Mike went to Melbourne. It could have been, but the author was lazy.

I find myself agreeing, albeit reluctantly, with a lot of the points that you are making. As limited as AIML may be, it is no more limited than much of what passes for intelligent behavior among human beings. The vast majority of our activities are habitual, and for many, the act of thinking is not an automatic response to common situations.

However it doesn’t alter the fact that most of the members of this forum want something more out of their software than AIML can provide. Honestly, who is going to be impressed by an artificial intelligence which is anything less than superhuman? For that, we need to harness a wide variety of techniques. Natural language processing isn’t enough by itself. Nor is speech recognition, machine vision or knowledge engineering.

Over the course of the last six decades, the study of artificial intelligence has diverged into many specialized areas, (almost) all of which are yielding practical and useful results throughout our lives already. Ultimately, a complete AI is likely to be the result of successful integration of (most of) these fields.

Thank you Andrew.

And if I produce a chatbot that uses AIML and is more advanced than the rest made by the members of this forum, would they continue to declare it is a dead end, that parsing grammar is the best? 

I once got distracted because some one suggested that dialogue management was what real computer conversation systems employed.  I studied things like Autotutor or Dipper, etc. I suspect eventually the forum’s discussion will turn towards that part of a collaborative interaction.  In the end it turned out to be just another way to classify and dissect communications, barely helpful to case manage areas with which to respond.

I mean there are typical lines of investigation this group is tracking. Eventually they will come back to content.  When the database includes millions of facts, when rules are captured for deductive reasoning, when inferences can extrapolate the “missing pieces” which seem to pop up in NLP perspectives, what is that but a very powerful search engine?  Are search engines intelligent?  Can they be wise? 

My system already contains a common sense database.  It can also tap into an ontology called Cyc.  I’m there already, and so what?  Cyc has reasoning.  I have some of that too and planning.  Fine.  I should just turn it loose to mine the internet - it could.  So where do you go after you’ve added all these desireable technologies which this group seeks?  What does a chat bot “think” about?

AIML persists because it includes the other aspects of chat bots.  It has addressed the psychology which seems to be avoided by the scientists here.  Where is the emotion in getting the train schedule right? Can any of these conversations support a chat bot’s attitude?

Gary Dubuque - Oct 24, 2010:

My system already contains a common sense database.  It can also tap into an ontology called Cyc.  I’m there already, and so what?  Cyc has reasoning.  I have some of that too and planning.  Fine.  I should just turn it loose to mine the internet - it could.

Where can we try out your software to evaluate its capabilities for ourselves?

The last release was at http://www.aimlpad.com, but I’m preparing an upgrade with a NLG extension so it can dynamically generate categories easier. It can also be found at SourceForge as Program-N, although that release is a little old.

The next evolution will be to align the personality parameters outlined in the character bible with the automatic language generation.  Thus the more outgoing personality might restate the obvious more often while the timid personality might interject “ums” and “ers” or the paranoid might use only third person, etc.  I’m hoping that the driving variables can be set during the conversation as the bot changes moods.  That way the automatic NLG can shift from talkative to reserved or maybe optimistic with mostly future tenses as the topics change.  Since AIMLpad has a wrapup api after the standard AIML response is generated, I’m looking to paraphrase some of the outputs, especially the multiple sentence input responses, with this tool.

In the spirit of AIML, AIMLpad is open source.  It was basically designed to help write AIML, but has been an experimental testbed for some of the themes in AI that are more focused towards conversations.

Gary I’ve been thinking a lot about your comments and the pros and cons of AIML and its fuzzy parsing approach using regular expressions, versus rigorous deep parsing and analysis using context free grammars. I’ll continue to explore the latter approach but in the meantime, I do have one more question for you.

Given your motivation for constructing a chatbot, have you considered making the leap to incorporating your efforts in a physical robot of some sort?

Andrew, I’ve only played with toys like Lego.  I do have a little robot that is driven by my pocket PC.  It has sensors to detect walls and stuff. The PPC is wireless so I could run the robot from my desktop.  I don’t play with it much anymore since the dogs would probably use it as a chew toy.

I could drag out my X10 things.  Once I had my whole house automated, but my wife thought it was haunted so I stashed all those toys away too.  I’ve been thinking I could do up the RV to have a mind of its own.  That would make vacations even more fun.

So did you explore (evaluate for yourself) AIMLpad?  It has got a draft of the NLG function included now - even though it has problems generating some kinds of questions.

Nathan, AIMLpad can store facts as kernel statements, that is, the triples used in Web 3.0, the semantic internet.  It might be a way to first try out some of your ideas before you get down to the specifics of building your design.

Thank you, Gary. I tried AIMLpad. This is the log of a chat.

Hello.
H E double hockey sticks oh!
Tom is a boy.
Are we still talking about Not many people express themselves that way?
Who is Tom?
Some famous celebrity, right?
What is Tom?
It is a good question. Have you tried surfing the web?

What should I do if I want AIMLpad to remember the fact that I just told it? Is there any existing scripts to support such functionality?