 |
|
|
 Okay, so what is GEN.I.E. 2.0 and why should you care? The GENeral Inference Engine is actually the 4th in a series of natural language processing programs (NLP) I've written over the years. For those of you not familiar, NLP describes computer/human interaction by means of normal language or chat sessions, much like you would see with instant messaging. In fact, these systems have been around for quite some time and with varying approaches (and success rates). I'll start you off with a little history: These first three programs I didn't write, nor do I have any association with their creators. You can go to their respective websites for details and history much better than I can do here.. Each one should give you an excellent feeling for the types of things that can be accomplished through NLP. Alice - Alice is what we call a chatterbot. Like an instant message chat session you type in normal speech and the computer approximates the most "human" response using pattern matching and a series of canned phrases. While this seems very intelligent to the person interacting wtih it, the rules for such an engine are actually quite simple in complexity, but numerous (over 3 million last I heard). ALICE has consistantly won NLP awards. Cyc - Cyc also works through a natural language interface but through a completely different mechanism. It uses millions of Boolean Logic rules and actually tries to calculate correlations between objects. For example it would know that the parent of your parent is a grandparent. Mindpixel - This site isn't up any more, and the founder is no longer with us. A Mindpixel was the idea that AI information could be boiled down into singular positive or negatives. True/False questions. While the site never really took off and had it's share of controversy , it did use one very interesting idea that I wish to note (and heap great amounts of credit for here), something far ahead of it's time. That idea being, using the internet as a source of corpus and a user community to build up it's heuristics. Nowadays this isn't an uncommon practice, but Mindpixel was the first as far as I know. Now for the programs I did write - MARGE 1995- This is a program I wrote as a sophmore in college. Marge is a chatterbot consisting of about 150 rules and 500 responses. Not the most powerful thing I've ever built, but stronger than the earliest chatterbots. I still have MARGE sitting on a disk somewhere written in SNOBOL of all things. FREDI 1997- Ah the memories. I wrote this program for my final project in AI class senior year in college. I think I still have it on a disk somewhere written in a thousand lines of garbled PASCAL. Fredi could recognize around 200 words and could perform some rudimentary actions. For example, if you typed "What is 6 plus 7?", Fredi could recognize the structure of that sentence and return "13". You could perform simple tasks like changing the background color of the screen, or feed it crude boolean rules that it could remember later. GEN.I.E. 1.0 1997- Another program I wrote in college. The idea behind the original GEN.I.E. was that I could feed the program a corpus of text and it would learn boolean rules from it. For example if the text contained "dogs ARE mammals" the program would pick up on the are/is/isn't/can etc. keywords and associate positive or negative correlations. While it generated lots of interesting rules/correlations, I never got around to putting in a chatterbot interface to it... until now. INTRODUCING GEN.I.E. 2.0 My idea for GEN.I.E. 2.0 is to make a "best of breed" web based NLP program. it's going to start out like MARGE, a simple chatterbot program that I'll most likely end up coding from my original MARGE code. My next step is going to "open up" the rules engine so that logged in users can add their own heuristics. Users will be able to rate GEN.I.E.'s responses so that only the best answers pop up. Of course, this means I'll have to put in safety features for people to report spammers/disrupters as the net isn' t always a friendly place. Step two is adding boolean logic. Like the original GEN.I.E. I'm going to key off from specific phrases (are/is/not etc.). Rule generation and inferences will be automatic. How deep the rules go and how complex they are...well, I'll see how popular the site is before I decide how much time to spend on it. Step 3 is to add functionality like FREDI on the highest level possible. Semantic network structure for searches actions or general logic. I'm not sure I'll ever get to this step. In fact, I'd be pretty happy with just a really good chatterbot. Like this idea? Want to contribute some time to the project? Maybe you just want to play around and see what GEN.I.E. can do for you... |
||||||||||||||