The features are combined in a log-linear way.

In Table 5 we can see that the performance of the heuristic phrase model can be increased by 0.6 BLEU on TEST by filtering the phrase table to contain the same phrases as the count model and reoptimizing the log-linear model weights.

Log-linear interpolation of the count model with the heuristic yields a further increase, showing an improvement of 1.3 BLEU on DEV and 1.4 BLEU on TEST over the baseline.

The translation process is implemented as a weighted log-linear combination of several models hm(e{, sf , ff) including the logarithm of the phrase probability in source-to-target as well as in target-to-source direction.

The log-linear interpolations pint(f|é) of the phrase translation probabilities are estimated as

As a generalization of the fixed interpolation of the two phrase tables we also experimented with adding the two trained phrase probabilities as additional features to the log-linear framework.

With good log-linear feature weights, feature-wise combination should perform at least as well as fixed interpolation.

Specifically, we modify the log-linear policy p(a|s; q, 6) by adding lookahead features gb(s, a, q) which complement the local features used in the previous model.

A Log-Linear Parameterization The policy

function used for action selection is defined as a log-linear distribution over actions: €9-¢(s,a)

where Pr(e| f) is the probability that e is the translation of the given source string f. To model the posterior probability Pr(e| f) , most of the state-of-the-art SMT systems utilize the log-linear model proposed by Och and Ney (2002), as follows,

In this paper, u denotes a log-linear model that has Mfixed features {h1(f,e), ..., hM(f,e)}, ,1 = {3.1, ..., AM} denotes the M parameters of u, and u(/1) denotes a SMT system based on u with parameters ,1.

In this paper, we use the term training set to emphasize the training of log-linear model.