P R(E|F) = P R(F|E)P R(E)P R(F)P R(F|E) = XP R(F, A|E)BUT IN THIS PAPE...

2003):

P r(e|f) = P r(f|e)P r(e)

P r(f)

P r(f|e) = X

P r(f, a|e)

But in this paper, we skip this step as we found out

a

the order of words in information need part is not

EM-algorithm is usually used to train the align-

an important factor. In our collected CQA archive,

ment models to estimate lexicon parameters p(f |e).

question title and information need pairs can be con-

In E-step, the counts for one sentence pair (f ,e)

sidered as a type of parallel corpus, which is used

are:

for estimating word-to-word translation probabili-

ties. More specifically, we estimated the IBM-4

P r(a|f, e) X

δ(f, f

_j

)δ(e, e

_a

_j

)

c(f |e; f, e) = X

model by GIZA++

⁴

with the question part as the

source language and information need part as the tar-

i,j

get language.

P r(a|f, e) = P r(f, a|e)/P r(a|e)

In the M-step, lexicon parameters become:

5 Experiments and Results

c(f |e; f

^(s)

, e

^(s)

)

p(f |e) ∝ X