Phramer is an open source statistical phrase-based MT decoder.
Why did you write Phramer?
The goals of the Phramer project were: write an extensible, source-code available (initially only for the authors), written in a nice programming language (w/o segmentation faults and non-scripting language), compatible with Pharaoh, that can be used in Statistical MT research.
Why would I prefer Phramer instead of Pharaoh?
a) License. Phramer is an Open Source program (BSD license), thus the limitations are minimal. Pharaoh is currently available under a much restrictive licensing
b) Source code availability. Phramer is Open Source, Pharaoh is currently not.
c) Cross-platform. Phramer is written in Java and it was tested in Linux and Windows 2000/XP. We think it should work on other platforms without any issues
d) Speed. Phramer is currently 1.5 - 3 times faster than Pharaoh 1.2.3 (not including data files loading time). We continuously work on improving the performance
e) Extensibility. Phramer has a modular architecture; a lot of hooks are provided
f) Features. Phramer has currently more features than Pharaoh 1.2.3 (not including undocumented features, that we are not aware of). Some of these features provide good ways to extend the decoder. See the list of additional parameters.
· Java 5/1.5 or higher
What's New in This Release:
· The newer version fixes some bugs, it has double the speed of the previous version (under regular conditions), it introduces multi-threading for both decoding and MER training, which takes advandage of multi-core and multi-processor machines, it also adds some new features (for decoding, MERT, framework). It also introduces benchmarking testset.