2.5 QDC POTENTIAL ABSOLUTE MINIMA OR MAXIMA THAT THE VALUES CAN ACHI...

3.2.5 QDC potential

absolute minima or maxima that the values can

achieve and their likely values. Furthermore, al-

We performed a manual examination of the 500

though these constraints are manually derived for

TREC2002 questions

²

to see for how many of these

our prototype system, they are fairly general for the

questions the QDC framework would apply. Being

human life-cycle and can be easily reused for other,

a manual process, these numbers provide an upper

similar questions, or for more complex dossiers, as

bound on how well we might expect a future auto-

described below.

matic process to work.

We noted that for 92 questions (18%) a non-

Birthdate

trivial constraint network of the above kinds would

apply. For a total of 454 questions (91%), a simple

Leonardo

Painting

reciprocal constraint could be generated. However,

for 61 of those, the reciprocal question was suffi-

Deathdate

ciently non-specific that the sought reciprocal an-

swer was unlikely to be found in a reasonably-sized

Figure 1. Constraint Network for Leonardo ex-

hit-list. For example, the reciprocal question to

ample. Dashed lines represent question-answer

“How did Mickey Mantle die?” would be “Who died

pairs, solid lines constraints between the answers.

of cancer?” However, we can imagine using other

facts in the dossier to craft the question, giving us

We also note that even though a constraint net-

“What famous baseball player (or Yankees player)

work might have been inspired by and centered

died of cancer?”, giving us a much better chance of

around a particular question, once the network is

success. For the simple reciprocation, though, sub-

established, any question employed in it could be the

tracting these doubtful instances leaves 79% of the

end-user question that triggers it.

questions appearing to be good candidates for QDC.

There exists the (general) problem of when more

than one set of answers satisfies our constraints.

4 Experimental Setup

Our approach is to combine the first-round scores of

the individual answers to provide a score for the