In this paper http://www.aclweb.org/anthology/D13-1176, in "2. Framework" it says

We begin by describing the modelling framework underlying RCTMs. An RCTM estimates the probability $P(f \ | \ e)$ of a target sentence $f = f_1, ..., f_m$ being a translation of a source sentence $ e = e_1, ..., e_k$. Let us denote by $f_{i\ : \ j}$ the substring of words $ f_i, ..., f_j $. Using the following identity,

$$ P(f \ | \ e) = \prod_{i=1}^m P(f_i \ | \ f_{1:i−1}, e) $$

an RCTM estimates $P(f|e)$ by directly computing for each target position $i$ the conditional probability $P (f_i|f_{1:i-1}, e)$ of the target word $f_i$ occurring in the translation at position $i$, given the preceding target words $f_{1:i-1}$ and the source sentence $e$.

I understand the identity that they used, but how is it implemented in terms of code? How can a module calculate such a probability?


For this specific paper, the conditional distribution is calculated as follows:

$ P(f_i = v |f_{1:i-1} ) = \frac{\exp(o_{i,v})}{\sum_{v=1}^V\exp(o_{i,v})} $

on page 1702

| improve this answer | |
  • 2
    $\begingroup$ You should edit the question? $\endgroup$ – HelloWorld Sep 6 '16 at 6:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.