‘able 2: Segmented Model Scores.

We find that when using a ood segmentation model , segmentation of the lorphologically complex target language im-roves model performance over an unsegmented aseline (the confidence scores come from boot-3rap resampling).

So, we ran the word-based baseline system, the segmented model (Unsup L—match), and the prediction model (CRF—LM) outputs, along with the reference translation through the supervised morphological analyzer Omorfi (Piri—nen and Listenmaa, 2007).

These are derived from the same unsupervised segmentation model used in other experiments.

performance of unsupervised segmentation for translation, our third baseline is a segmented translation model based on a supervised segmentation model (called Sup), using the hand-built Omorfi morphological analyzer (Pirinen and Lis-tenmaa, 2007), which provided slightly higher BLEU scores than the word-based baseline.

We therefore trained several different segmentation models , considering factors of granularity, coverage, and source-target symmetry.

The goal of this experiment was to control the segmented model’s tendency to overfit by rewarding it for using correct whole-word forms.

We also build in parallel a segmentation model to select 3,- from the set {new}, same}.

We build two segmentation models , one trained on contributions of less than four tokens, and another trained on contributions of four or more tokens, to distinguish between characteristics of contentful and non-contentful contributions.

No segmentation model is used and no ILP constraints are enforced.