This is a library for structured prediction. It was primarily developed by Taylor Berg-Kirkpatrick. Other contributors include John DeNero, Aria Haghighi, Dan Klein, Jonathan Kummerfeld, and Adam Pauls.

To use the library, download it one of these ways, and include it in your code as described below:

• Download .zip
• Download .tar.gz
• git clone https://github.com/tberg12/murphy.git

This library was initially released as supplementary material for the experiments described in:

An Empirical Analysis of Optimization for Max-Margin NLP, Jonathan K. Kummerfeld, Taylor Berg-Kirkpatrick and Dan Klein, EMNLP 2015

You will need to implement the following:

• Code that creates a LossAugmentedLearner, e.g. PrimalSubgradientSVMLearner, and calls train. This is your main interface to the learning code. At creation you set parameters such as the learning rate and regularization, and when you call train you provide the data, initial weights, and number of iterations.
• A class that extends LossAugmentedLinearModel. This is the interface through which the learning code calls your inference procedure. Given an instance, your code will find the best structure (b) under your model with loss-augmentation relative to the gold (g). You will return an UpdateBundle that includes the loss of b relative to g, and the features active in b and g.
• [if using sparse updates] Modifications to your inference procedure to use getCount on a LazyAdaGradResult to get weights. This handles the delayed updates (see the paper above for further informtation).

Then include this code on your classpath, compile, and you're ready to go!

For an exmple, see the files in the example directory. Note, the classification it does is trivial, but it shows how you put the pieces together. You can build and run the example with build_and_run_example.sh (assuming you have javac, java, and standard command line tools - find, echo, mkdir).

The library comes with three versions of the structured predicition code:

• structpred, the basic implementation
• floatstructpred, a variant that uses floats instead of doubles to save space and time
• lazystructpred, a variant that allows for sparse updates

Note that the dual methods (NSlackSVMLearner) are only implemented for the basic version.