This is in the form of a set of e-mail responses.  I have only the
ones written by TDL, which in turn excerpt what I wrote.  That's
'cause I save old e-mail, but not stuff I send :-) This entire
exchange took place in early April of 2002.

This describes a very interesting framework for object descriptions.
It may or may not be exactly what Phantasmal winds up implementing.
It is, in any case, an excellent thought exercise.

==========================================================================
> I've worked out a framework for how the mud can
> describe objects and
> groups of objects.  It formalizes not only how the
> mud describes objects,
> but how the player can refer to objects.  It also
> formalizes when and how
> definite articles are used.
> 
> I'm assuming that there's a class hierarchy of
> objects, featuring
> arbitrary multiple inheritance, and that an object
> can be an instance of
> multiple classes if it wants.  (I'm also assuming
> that inheritance is
> "virtual" in the C++ sense, and that subclases can
> selectively 
> "uninherit" things if they want to.  Neither of
> these are critical for the
> current discussion, though.)
> 
> I'm also only considering English for now.  Forgive
> me.
> 
> For simplicity, let's assume that an object is fully
> defined by
> the classes which it inherits.  That is, we'll only
> consider descriptions
> which are based on an object's classes, and not its
> attributes or status.
> We'll see how robust a system we can get with that
> before we get confused
> by attributes.
> 
> Every class, then, has to provide the following:
> 
>    - A set of nouns by which individual instances of
> this class
>      may be referred.  These may be compound nouns
> ("end table",
>      "remote control"), but they may not contain
> articles.  For
>      every noun that you want to add to this list,
> make sure it
>      sounds right to say "the", "one", and "a" (or
> "an") in
>      front of it.
> 
>    - A set of plural nouns by which multiple
> instances of
>      this class may be referred.  Examples are
> "remote
>      controls" and "geese".  For every noun you want
> to add to
>      this list, make sure it sounds right to say "a
> couple",
>      "several", "five" , and "many" in front of it.
> 
>    - A set of collective nouns by which collections
> of instances
>      of this class may be referred.  This is things
> like
>      "livestock", "equipment", "silverware", or
> "weaponry".  For
>      every noun you might add in this list, make
> sure it sounds
>      right to say "some" and "a fair amount of" in
> front of it.
> 
> The latter two categories are not the same.  Plural
> nouns described some
> "number" of objects, while collective nouns describe
> some "amount" of
> something.  The "amount" in this case is discrete
> (since I'm not dealing
> with stuff like "5.3 ounces of water"), but it's
> still grammatically
> distinct from simple plurality.
> 
> (None of these categories handle proper nouns like
> "Bob."  Ignore those
> for now.)
> 
> Okay.  Nothing profound so far.  It's probably what
> you were planning,
> anyway.  But I'm trying to formalize this stuff.
> 
> Given the above, we can now implement several
> low-level functions.  It
> should be obvious how to implement each.
> 
>  bool doesDescribeOne(class *o, string s);  //
> returns true if the given
>                                             //
> string describes a single
>                                             //
> instance of the given class
> 
>  bool doesDescribeMany(class *o, string s); //
> returns true if the given
>                                             //
> string describes several
>                                             //
> instances of the given class
> 
>  string describeOne(class *o);   // returns a string
> describing one
>                                  // instance of the
> given class, not
>                                  // including any
> articles.  It either
>                                  // chooses a noun
> randomly, or (more
>                                  // likely) chooses
> the one marked
>                                  // "preferred".
> 
>  string describeMulti(class *o, int num);
>                                  // returns a string
> describing several
>                                  // instances of the
> given class.  It 
>                                  // can choose
> whether to use a
>                                  // plural noun or a
> collective noun.
>                                  // Based on the
> given number, it may
>                                  // prepend a
> specific number ("five") or
>                                  // a general
> quantifier ("several").
> 
> Easy enough.  But next comes the kicker.
> 
>  string describeObjects(vector<object *> objects,
> int maxPhrases);
> 
> This function takes a set of specific objects and
> returns a string
> describing that set of objects, including articles,
> conjunctions,
> and all.  The "brevity" of the description is
> controlled by the parameter
> "maxPhrases".
> 
> Suppose it's given a spoon, a fork, a knife, a
> spatula, and a banana.
> 
> If maxPhrases is 1, the best it can to is "several
> objects" or "some
> stuff".  (It finds the class common to all objects,
> which in this case is
> "object", and calls describeMulti() on that class.)
> 
> If maxPhrases is 2, it can do "some kitchenware and
> a banana".  (It
> notices that the silverware and the spatula have a
> common class, and so it
> groups them accordingly.)
> 
> If maxPhrases is 3, it can do "some utencils, a
> spatula, and a
> banana".  (It took the opportunity to differentiate
> the utencils from the
> spatula.)
> 
> You get the idea.
> 
> Internally, this function takes the set of objects,
> finds the places they
> appear in the class hierarchy, and tries to find a
> set of "common
> ancestor" classes that covers all objects.  It's an
> optimization
> problem: you want the number of ancestors to get as
> close to maxPhrases as
> you can without going over, and you want to minimize
> the distance from the
> objects to the ancestors you choose.  (Otherwise
> you'd always say "an
> object, an object, an object, and some objects.")
> 
> I'm not saying that this optimization is easy, but
> it _is_ self-contained,
> and it's easy enough to implement a dumb version
> that will work in many
> cases.  This can be made as smart-and-slow or
> dumb-and-fast as
> desired, without affecting the design of the rest of
> the system.
> 
> One part of this function's strategy isn't self
> contained: the choice of
> whether to use definite or indefinite articles for
> each phrase it
> returns.  That choice depends on "conversation
> context".  I believe I
> have a strong sense of how that choice should be
> made, but I'll
> ramble about that some other time if you like.
> 
> This function doesn't solve all the world's
> problems, but I believe it's
> an important low-level tool.  Higher level code will
> choose what objects
> to describe, which to leave out, and where it wants
> the descriptions to
> show up.  Once the higher level code has decided it
> wants to describe a
> set of objects somewhere in a sentence, it calls
> describeObjects() to get
> that description, all bundled up and grammatically
> happy.
> 
> If this is still in the realm of "yeah, yeah,
> someday" to you, forgive me
> for making you read all that.  But I feel like I
> have a concrete sense
> now of how to get this stuff working for real, so I
> wanted to share.
> 
> (I've left lots of stuff out, but this is already
> way too long.  I'll shut
> up for now.)
> 
> --Tom

==========================================================================

> > More of it is doable right now than you realize
> 
> Not sure about that.  I sent it to you because I
> believe just about 
> all of it is doable.  :)
> 
> >   It *does* leave out explicit adjectives
> 
> Yes, and to get those right, the "search" that tries
> to describe a set of
> objects in maxPhrases phrases needs to consider
> grouping based on
> adjectives as well.
> 
> Suppose describeObjects() is given a red pen, a red
> pen, a red pen, a
> green pen, and a banana.  Let's assume that all the
> pens are just pens,
> and the color is an attribute (as opposed to a
> class).
> 
> If maxPhrases is 1, the best we can do is "some
> stuff" or "some oblong
> objects".
> 
> If maxPhrases is 2, we can do "some pens and a
> banana."
> 
> If maxPhrases is 3, we can do "some red pens, a
> green pen, and a banana."
> 
> This last bit means that the "optimization" needs to
> consider adjectives
> as well.  It's not necessarily straightforward, but
> I think it's doable.
> 
> Incidentally, the choice for which attributes to
> categorize on first
> should be class-specific.  If we're asked to
> describe several crayons,
> it's most natural to differentiate based on color
> first (then possibly
> sharpness, etc).  If we're asked to describe several
> apples, we should
> differentiate by size, then possibly ripeness.
> 
> > the boolean query for
> > whether a string describes the object combined
> with
> > the ability to describe a set of them handles
> that.
> 
> For recognizing adjectives that a user has typed,
> yes, that's true.
> 
> > The simplest is that it doesn't
> > allow an object described individually to have
> > variable levels of detail.
> 
> That would be cool.  An extra bonus.
> 
> >   Also, do "doesDescribeOne" and its relatives
> parse
> > and accept articles?
> 
> I'm assuming no, that they're supposed to deal with
> noun phrases
> sans-article.  describeObjects(), being a bigger and
> stompier function,
> does have to deal with them.
> 
> Or to put it another way: it's okay if
> doesDescribeOne() and kin handle
> articles, as long as we have variants that handle
> the un-articled
> nouns.  This is important because some higher-level
> code (such as
> describeObjects()) will need to do complex article
> determination itself,
> and so we don't want the very low level code
> (doesDescribeOne(),
> etc) always imposing its own articles.
> 
> However, my reasoning for that is based on my
> assumptions for how definite
> and indefinite articles should work.  And I don't
> have time to get into
> that this morning.  :)
> 
> --Tom

==========================================================================

> >   This framework also fails to address the
> question of
> > multiple possible groupings for a single object.
> > ...
> >   Though I suppose that's really what parent
> classes
> > are *for*...  "The doorman" could be a parent
> class.
> 
> That's what I'm assuming.
> 
> I should take a couple minutes to explain my take on
> definite vs
> indefinite articles.  Here's what I believe:
> 
> Objects should never be tagged as "definite" or
> "indefinite".  Definiteness
> (definity?) is never a property of a particular
> object; it is a property
> of a phrase, and which object(s) that phrase was
> most recently used to
> describe.
> 
> For example, suppose I have a red pen, a blue pen,
> and a green pen.
> Suppose I point to the red pen and say "that is a
> red pen."  At this point
> in the conversation, the phrase "the pen" maps
> unambiguously to the red
> pen.  So does the phrase "the red pen."
> 
> (Ignore how we'd implement this.  Just see if you
> agree with my logic.)
> 
> Suppose I then point to the green pen and say "that
> is a green pen."  Now
> the phrase "the pen" maps to the green pen, as does
> "the green pen."  But
> the phrase "the red pen" still maps to the red pen.
> 
> So it's never the case that an object is "definite"
> or "indefinite".  It's
> only definite or indefinite with respect to a
> particular potential
> description of the object.
> 
> For this reason, I don't think of it as a property
> of the object at
> all.  I think of it as a property of the phrase. 
> Each phrase ("pen",
> "red pen", "pens", "warm-colored pens") has a
> pointer to the object (or 
> set of objects) which it most recently described. 
> (Often it's null.)
> 
> The mapping from a phrase to its
> most-recent-referents must be updated
> whenever _any_ object which matches the phrase is
> mentioned.  For example,
> If "the pens" maps to all three pens, but then
> someone says "the
> winter-colored pens", we have to udpate the phrase
> "the pens" as well as
> the phrase "the winter-colored pens".
> 
> This makes sense to me.  Does this make sense to
> you?
> 
> Now let's consider describeObjects(), which is given
> a set of objects and
> a limit (maxPhrases) on the number of phrases it may
> use.
> 
> The job of describeObjects() is to choose up to
> maxPhrases phrases.  Each
> phrase will describe a subset of the objects.  The
> selection of these
> phrases is a partitioning process: it decides how to
> separate the given
> objects into groups, where all objects in a group
> share the same
> description.
> 
> Once it has decided to describe a subset of the
> objects by a given phrase
> ("utencils" or "leaky red pens" or something), NOW
> is the time for it to
> choose definite or indefinite.  It queries the
> phrase to ask if that
> phrase was most recently used to describe this
> particular set of
> objects.  If it was, describeObjects() uses the
> definite form for that
> phrase  If it wasn't, describeObjects() uses the
> indefinite form for that
> phrase.  (In the latter case, it then updates the
> phrase so that it will
> be definite for these objects _next_ time.)
> 
> --Tom

==========================================================================

> >   This doesn't address the same objects that would
> be
> > tagged as definite under the old scheme:  "the
> floor",
> > "the sky".
> 
> Yes, I'm convinced that those are abberations in the
> language.  An
> artifact of living on the earth is that you know, no
> matter where you
> are, there's probably a floor.  And notably, it
> doesn't show up in
> descriptions _specifically_ because you expect
> everyone to assume it's
> there.
> 
> This blanket use of "the" for floor works because
> there's pretty much
> always exactly one floor in any given place.  And
> you can imagine
> exceptions:
> 
>     You're in a nightclub.  Ugly people are wiggling
> on a shiny
>     dance floor.
>     > drop dead rat on floor
>     Do you mean the floor, or the dance floor?
> 
> One way to get this to work: whenever you move to a
> new room, it
> checks to see if there's a single obvious "floor"
> object (which there
> almost always is).  If there is, it sets the
> most-recent-referent of 
> the phrase "floor" to that object.  If there isn't,
> it sets the
> most-recent-referent of the phrase "floor" to null.
> 
> > > look at my Tori Amos CD
> > You stare lovingly for a moment at the cover
> before
> > realizing that, cute or no, you don't think much
> of
> > her music.
> > ...
> >   Note that funky articles in a couple of places. 
> > It's not necessarily a problem, but it *is* a
> little
> > odd.
> 
> It is sort of odd, and I'm open to improvements.
> 
> Incidentally, I don't think we should ultimately be
> ignoring articles
> in user commands.  As in an earlier example, "take a
> pen" and "take the
> pen" mean two completely different things.  (The
> first implies some
> freedom on the part of the recipient, while the
> latter implies
> prior agreement on a particular pen.)
> 
> If you think about it, text game pidgin english
> commands like "take
> pen" or "open door with blue key" almost always have
> an implied "the" when
> articles are left out.  I'm tempted to make that
> explicit: articles are
> supported, and if you leave out an article, it
> assumes the definite
> article.
> 
> --Tom

==========================================================================

> > in the example above using the phrase "the cover"
> when
> > "my Tori Amos CD" was the actual phrase used.
> 
> I've been working through this too.  If you walk
> into a room and see a
> box, you should be able to say "look at the front of
> the box" even if the
> "front" has never been mentioned.  (Same for "the
> cover of the cd".)  This
> is similar to "the floor".  You assume all boxes
> have fronts and all cd's
> have covers.  So you're right, some objects and
> parts-of-objects can be
> definite-articled for free, but only in very limited
> ways.  I don't claim
> to understand all the implications, but I believe
> strongly that the answer
> is _not_ to tag the objects themselves as
> "definite".
> 
> Actually, maybe I _do_ know what's going on.  As I
> said before, objects
> aren't definite, but _phrases_ may be definite for
> particular objects.  In
> the case of prepositional phrases, those are
> definite only with respect to
> the prepositional object (or whatever it's called).
> 
>    There's a box here.
>    > look at the front of the box
> 
> It first disambiguates "the box", then sees if the
> phrase "front of <that 
> particular box>" has a most-recent-referent.  If it
> doesn't, the game
> "searches" to see if there's exactly one reasonable
> referent.  If so, it
> lets you get away with the definite article.  The
> same scheme would work
> for the floor.
> 
> So this changes the strategy slightly.  When the
> player uses an indefinite
> article, the game feels free to choose anything that
> matches the
> description.  When the player uses a definite
> article, the game
> first checks to see if that phrase is bound to a
> most-recent-referent.  If
> so, it uess that.  If it doesn't, it searches to see
> if there's a single
> obvious referent.
> 
> > Even if you've only just seen him, he really isn't
> just *a* King of
> > Spain.
> 
> Hehehehe.
> 
> --Tom

==========================================================================

> > > look up
> > You see the Berkeley Naked Guy!
> 
> Hanging on the ceiling above you?  Creepy!  And
> unsanitary!  :)
> 
> > Presumably the King of Spain is also the most
> recent
> > "King of Spain" in cache as well, at least
> > conceptually.  So the "distinctiveness" tag is
> really
> > sort of a probability of that object being
> considered
> > "permanently in cache"
> 
> That makes sense.
> 
> --Tom

==========================================================================

> > if instead of the player
> > saying "the cover of the CD" they say "the cover",
> we
> > should still check the CD.
> 
> Yup.  That would be good.
> 
> > You're in the kitchen.  Again.  If only you could
> > avoid the need for food.  And the hunchback in
> > lederhosen.
> 
> ROFL
> 
> > > look at stem
> > (You look at the stem of the apple in the
> fruitbasket)
> 
> I would argue that here, unless there's exactly one
> object with a stem, it
> should say "which stem?".  But that's because I read
> "look at stem" as
> "look at the stem," which is ambiguous.
> 
> > the nappy and suedelike leather of these
> > DeutscheBritches.
> 
> :)
> 
> >   Hm.  Okay.  That's certainly fair.  I'm trying
> to
> > figure out if any detail of a definite-articled
> object
> > is also definite-articled.
> 
> I don't think so.  While bob's head is part of bob
> (and so we think of the
> head as being subordinate), in the sentence "look at
> bob's head", bob
> isn't the direct object.  The head is.  The
> bob-possessive serves,
> basically, as an adjective modifying "head".  I
> believe the the
> definite-ness of the detail is independent from the
> definite-ness of the
> thing-which-the-detail-happens-to-be-on.
> 
> To pick an extreme example:
> 
>    > touch the front of a box
>    You pick and box and touch its front.  Wheee!
>    > touch the front of the box
>    You touch the front of the same box.  It's less
> exhilerating the
>    second time.
> 
> In both cases, the prepositional phrase is
> disambiguated (so we know what
> it's "of"), then we disambiguate "front" based on
> that.
> 
> > > It first disambiguates "the box", then sees if
> the
> > > phrase "front of <that 
> > > particular box>" has a most-recent-referent.
> 
> Oh.  I've sort of said that already.  Oops. :)
> 
> --Tom

==========================================================================

> > To pick an extreme example:
> 
> Here's an even more extreme example, which I don't
> claim we should
> support, but it clarifies my delusions about what
> definite articles do:
> 
>     You're in a tavern.  There are several tables,
> each with
>     several glasses.
> 
>     > look at the glass on the table
> 
>     Which table?
> 
>     [ it failed to disambiguate the prepositional
> phrase, so it gave up.
>       it didn't even bother trying to disambiguate
> the noun phrase that
>       used the prepositional phrase. ]
> 
>     > look at the glass on the first table
> 
>     Which glass?
> 
>     [ it successfully disambiguated the
> prepositional phrase, but
>       when it tried to disambiguate "the glass" on
> that table, it
>       failed. ]
> 
>     > look at a glass on the first table
> 
>     It's dirty.
> 
>     [ it successfully disambiguated the prepositions
> phrase, then
>       chose a glass from that table. ]
> 
>     > look at a glass on a table
> 
>     It's dirty.
> 
>     [ It chose a table, then chose a glass on that
> table.  In
>       this case, it would have to backtrack and
> choose another table
>       if the first choice had no glasses. ]
> 
>     > look at a glass on the table
> 
>     It's dirty.
>     [ It disambiguated the prepositional phrase
> (coming up with the
>       same table as last time), then chose a glass
> from it. ]
> 
> Those all make sense to me.  The remaining case is
> 
>     > look at the glass on a table
> 
> By my claims, it should disambiguate the
> prepositional phrase by picking
> one (randomly?), then try to disambiguate "the
> glass" on that particular
> table.  If there is exactly one glass on that table,
> we're in luck, and it
> chooses it.  If there's more than one glass on that
> table, but we referred
> to one of them recently, it chooses that one. 
> Otherwise, it says "Which
> glass?"
> 
> This seems like the right behavior, though.
> 
> --Tom

==========================================================================

> >  But that's because I read "look at stem" as
> > "look at the stem," which is ambiguous.
> 
> Incidentally, I'm not adamant that "look at stem"
> means "look at THE
> stem."  If you'd prefer that it implies some
> imaginary article which is
> neither "the" nor "a", and which means "find the
> definitive interpretation 
> if there is one, and if not, pick one randomly", I
> wouldn't object.  My
> main obsession is getting "the" and "a", as the game
> uses them, as correct
> as possible.
> 
> And, as always, forgive me if I'm being pushy.  My
> repeated 
> explanations are for clarification, and aren't
> intended to force my
> uneducated viewpoint on you.  :)
> 
> --Tom

==========================================================================

> > Here's an even more extreme example, which I don't
> claim we should
> > support, but it clarifies my delusions about what
> definite articles do:
> 
> The more I think about it, the more I think this
> case will just work, if
> we disambiguate prepositional phrases before
> disambiguating the noun
> phrases that use them.
> 
> If we do try to do this stuff, here's a good test:
> 
>      You're in a tavern.  There are several tables,
> each with
>      several glasses.
>  
>      > look at a glass on the first table
> 
>      It's tall and skinny.
> 
>      [ it disambiguates the table, then picks a
> glass on it. ]
> 
>      > look at a glass on the second table.
> 
>      It's short and cracked.
> 
>      [ same deal ]
> 
>      > look at the glass on the first table.
> 
>      It's tall and skinny.
> 
>      [ this should come up with the same glass as
> before, since
>        "the glass" should disambiguate relative to
> the disambiguation
>        of the prepositional phrase. ]
> 
>      > look at the glass on the second table.
> 
>      It's short and cracked.
> 
>      [ same deal ]
> 
> I'm not sure how exactly to implement this, or
> exactly how to represent
>
the-mapping-from-context-specific-phrases-to-most-recent-referents.
>  Maybe 
> it won't be practical.  It's good to keep in mind,
> though.
> 
> --Tom

==========================================================================