We utilize minimum error rate training (MERT) (Och, 2003) to optimize feature weights of the paraphrasing model according to NDCG.

MERT is used to optimize feature weights of our linear-formed paraphrasing model.

S X3184 = arg min{Z ETTaDz-Label, 62,-; A3184, M” i=1 The objective of MERT is to find the optimal feature weight vector Xi” that minimizes the error criterion Err according to the NDCG scores of top-l paraphrase candidates.

Figure 2: Visual and acoustic feature weights .

To determine the role played by each of the visual and acoustic features, we compare the feature weights assigned by the learning algorithm, as shown in Figure 2.

Next we extract phrase pairs, Hiero rules and tree-to-string rules from the original word alignment and the improved word alignment, and tune all the feature weights on the tuning set.

The feature weights can be tuned on a tuning set in a log-linear model along with other usual features/costs, including language model scores, bi-direction translation probabilities, etc.

First, in addition to the preprocessing of training data and inserting recovered empty categories, we implement sparse features to further boost the performance, and tune the feature weights directly towards maximizing the machine translation metric.