[MUD-Dev] re: Language Parsing for NPCs

Aaron Mulder ammulder at alumni.princeton.edu
Mon Jun 18 17:55:20 CEST 2001


On Sun, 17 Jun 2001, Robert Zubek wrote:

> cool! i'm currently beginning ph.d. work in a similar area -
> understanding what people do in social multiplayer games. i
> suppose this isn't as language-intensive as your interest, but i'm
> butting against the nlp problem all the time as well (and what a
> nasty problem it is! :)

> i'm in the same boat, actually - i've a set of older logs, but
> they all involve one of my characters. :) if there were some way
> to get third-person logs of conversations without having to start
> my own mud, that would be truly wonderful. :) if you or anyone
> else know of a way to get such logs, please post!  (also, aaron,
> if you'd be interested in swapping logs, let me know...)

> and on a related note - having a standard corpus of mud
> conversation logs would be of *enormous* help to others trying to
> do research-quality work on natural language understanding for
> games. however, to my knowledge, no such standard corpora exist
> anywhere.


First, I strongly agree that a repository of generic conversational
logs would be really helpful.  But someone else pointed out that in
many environments, players have been trained to use keywords, so the
logs aren't as distinctive as they might be.

As for where I'm heading with all this, here's my theory so far:

  Step one is to process the sentence, identifying parts of speech
  and so on.  The problem is, this doesn't get you very far.  For
  example, "I ran a MUD with a really smart NPC."  Does "with a
  really smart NPC" refer to the MUD, or to the running of the MUD
  (as in, "I ran a MUD with a P3 and a T1.")?

  Step two is to map the words to concepts.  The is, we must
  understand (or at least acknowledge) the concepts of running
  software, of MUDs, and of NPCs.

  Step three is to relate all the concepts.  Thus we can determine
  that while a computer might be a platform on which one runs a MUD,
  an NPC isn't a tool you use to run a MUD.  On the other hand, an
  NPC would quite reasonably be associated with the MUD that is
  running.  So the relation of concepts is what we use to resolve
  the ambiguity in the sentences.

  Step four is to somehow convey the meaning of the sentence to the
  AI for the NPC.  I assume that the AI has access to the same bag
  of concepts, but we still need to somehow classify the sentence as
  a question, an order, etc.  The AI has to be able to figure out
  how to respond.  I haven't really made much progress in this area
  yet.

So then we have some other issues, such as handling specific nouns.
The NPC has to have senses and a memory, so that they will notice
their surroundings, and if they are asked about a specific item,
place, etc. they can understand.  As well, the NPC has to track the
preceding parts of the conversation, so that ongoing references make
sense.  I suppose this amounts to both short term and long term
memory.

I haven't gone very far here either, other than the thought that
everything the NPC experiences should be associated back to concepts
in the NPC's bucket o' concepts.  And somehow this has to work
without coding specific handling for every item - perhaps some sort
of keyword association within descriptions of things.


Anyway, that's the 20 cent tour.

Aaron

_______________________________________________
MUD-Dev mailing list
MUD-Dev at kanga.nu
https://www.kanga.nu/lists/listinfo/mud-dev



More information about the mud-dev-archive mailing list