Well, this is my first post here, so I thought I might start introducing myself and my current project first.

For a few years, I've been busy with both studying and working in the field of computational linguistics, and since last year's summer, I've been working on some kind of agent that will (hopefully) be one day able to actually "understand" natural language.

I used to chat alot while on IRC, so I thought implementing an agent that will be able to accompany me would be fun, but since then, the system's purpose is more one of being able to act semi-autonomous to help me with common tasks, like searching a medium (IRC, WWW, etc.), data mining the results, locating people and so on...

It's knowledge of words was acquired by parsing months of chat-logs, building a corpus of about 70.000 words, which was manually annotated to build a statistical POS-tagger later on (based on markov models.)

There are several sub-systems that ensure that the input is actually in a form that guarantees maximum usability to the NLP parser.
These include an algorithm based on the idea of double metaphone (heavily modified for my native language) to reduce orthographic errors, a language detection system based on an implementation of naive bayes, to allow to distinguish between input in several languages, thus resulting in a convenient way to sort out input which cannot be processed anyway.

After input data has been preprocessed this way, it is passed to the tagger, which identifies the class and syntactic meaning of a given word, i.e. verbs (predicate), nouns (subject / object) etc., which is for example needed to decide wether semantic statements like the following (common) example are actually the same:

"Thought is nonverbal speech" <---> "Speech is verbalized thought"

While it seems to be clear up to now that parsing NL seems to work fine in terms of feasibility (there is plenty of room for improvement, though), the impression of artifical intelligence can only be conjured with the presence of knowledge, and even more important, with the inference of knowledge.

I'm currently working on an implementation of what I call KREBLOP, which is an acronym for knowledge representation based on logic programming, which in turn is a smart-ass way of referring to an interpreted language based on Prolog, with the only purpose of making facts and according rules available to the NLP engine.
In this area, progress is somewhat lacking recently, so I'll guess I have to invest my next spare time right there...

Hmm, although I haven't described any implementation details, or even mentioned all components, I hope I was able to give an (maybe even interesting) insight of how a modern ELIZA bot could look under the hut...

If (or as soon as) speech recognition advances, I would like to port my efforts to some kind of robot system, which might evolve into a useful companion when it comes to housework...or maybe this only reflects my deepest wishes to command someone to clean the dishes...

Maybe even one of the readers is interested in this subject, or even better, has tried for himself to create such a bot.
Just don't hesitate to reply or contact me...