3).THE OBJECTIVE OF OUR WORK WAS TO PRESENT WHATTHE FEATURES...

section 2.3).

The objective of our work was to present what

The features mentioned above determined a space

we believe is a valuable conceptual framework;

Ψ; An answer a, in such feature space, assumed

more advance machine learning and summariza-

the vectorial form:

tion techniques would most likely improve the per-

Ψ a = ( ϑ, ς, $, % )

formances.

The remaining of this paper is organized as fol-

Following the intuition that chosen best answers

lows. In the next section Quality, Coverage, Rel-

(a ? ) carry high quality information, we used su-

evance and Novelty measures are presented; we

pervised ML techniques to predict the probability

explain how they were calculated and combined

of a to have been selected as a best answer a ? . We

to generate a final summary of all answers to a

trained a Linear Regression classifier to learn the

question. Experiments are illustrated in Section

weight vector W = (w 1 , w 2 , w 3 , w 4 ) that would

3, where we give evidence of the effectiveness of

combine the above feature. Supervision was given

our method. We list related work in Section 5, dis-

in the form of a training set T r Q of labeled pairs

cuss possible alternative approaches in Section 4

defined as:

and provide our conclusions in Section 6.

T r Q = {h Ψ a , isbest a i}

2 The summarization framework

isbest a was a boolean label indicating whether a

2.1 Quality as a ranking problem

was an a ? answer; the training set size was de-

Quality assessing of information available on So-

termined experimentally and will be discussed in

cial Media had been studied before mainly as a