File class1.doc
[extracted from class_a_crit2.txt, msg frm B. Moore dated 5/10/90]

		DEFINITION OF CLASS A QUERIES


Class A queries will be identified by exception.  Class A queries will
be those that are none of the following:


1. context dependent

There seem to be two broad subcategories here:

	a. queries containing explicit reference to a preceeding
	   answer or question, such as "What classes of service are
	   available on those flights?"

	b. queries whose scope is implicitly assumed to be limited
	   by a preceeding answer or question, such as "Which flights
	   go to Dulles?" in a context that limits attention to
	   some particular set of flights to Washington DC.

It is noted that some queries in the second subcategory could, in
isolation, also receive a reasonable context-independent
interpretation. For example, in context, "Please list an
interpretation of the classes," is likely to mean the classes
displayed in the preceeding answer, and thus is context dependent.  It
also could have a reasonable use referring to all classes.  Such
queries will be specially marked in the process of selecting class A
queries.


2. vague, ambiguous, disambiguated only by context, or otherwise
failing to yield a single cannonical database answer.

Some of the particular cases noted so far include:

	a. attachment ambiguities.  These will be excluded ONLY if
	   it is not possible for an ordinary person to pick the
	   preferred reading (without resort to context).

	b. "What does X mean?"  These are out, unless X is an
	   abreviation code that has a table that expands the code
	   into a descriptive word, phrase, or set of attributes.  The
	   query is not acceptable, however, if X is a code that has
	   more than one possible meaning according to what field it
	   appears in, unless there is disambiguating context WITHIN
	   the query, such as "What does fare code X mean?" 

	c. "Give me information about X."  These queries could be
	   allowed, if someone will produce a table of allowable
	   instances of X together with what information should
	   be provided.  Pending that happening, these queries
	   are out.


3. grossly ill-formed.

As long as the query is interpretable, only utterances that appear not
to be attempts to speak normal conversational English will be
excluded.  For example, we should exclude attempts to speak some
imagined form of "computerese" rather than normal English: "Origin
Dallas, destination Boston, list flights."


4. other unanswerable queries

Some subcases:

	a. queries not given a database answer by the wizard.  This
	   may include some queries that pass all our tests, but if
	   the wizard did not generate a DB query, then we don't
	   have anything to evaluate on.

	b. utterances that cannot be interpreted as queries, or that
	   are incoherent.

	c. queries that request information not in the database.

	d. queries that refer to the way that information is
	   presented.  

	e. "meta queries" about system capabilites or structure or
	   limits of the database.


5. queries from a noncooperative subject

Utterances that are clearly designed to try to break the system should
be excluded: "Given that city A is Oakland and city B is Fort Worth
show me all flights from A to B."


NOTES:

Minor syntactic or semantic ill-formedness -- if the query is
interpretable, it will be accepted, unless it is so ill-formed that it
is clear that it is not intended to be normal conversational English.

Presupposition failures -- all presuppositions about the number of
answers (either existence or uniqueness) will be ignored.  These are
the only types of presupposition failures noted to date.  Any other
types of presupposition failure that make the query truly unanswerable
will presumably result in the wizard being unable to generate a
database query, and will be ruled out on those grounds.

Multi-sentence utterances -- These will not automatically
be ruled out.  The examples cited so far are clearly
interpretable as expressing multiple constraints that
can be combined into a single query.


PROCEDURE FOR CLASSIFYING SENTENCES:

There are five general categories of non-class-A utterances, with an
important special subcase of context-dependent utterances.  We will
therefore use the following code:

        C  -- Hopelessly context dependent; COULD NOT reasonably be
              uttered with an unambiguous context-independent reading.

        C1 -- Context dependent, but COULD reasonably be uttered with
              an unambiguous context-independent reading.

	V  -- vague, ambiguous, etc.

	I  -- ill-formed (grossly).

	U  -- unanswerable (for other reasons).

	N  -- noncooperative subject.

NOTE: There are some context-dependent queries that could be forced to
have a context-independent interpretation, but it would be
unreasonable to do so, because of the large amount of data that would
be retrieved; for example, "Where are connections made?"  Such queries
will be classified as C rather than C1.