Version 1.0.0

This is a C++ 11 implementation of A Unified Bayesian Model of Scripts, Frames and Language (Ferraro and Van Durme, 2016). It has been developed and tested on Linux x86_64, under G++ >= 4.8.4. It should also compile on a Mac, though a couple Makefile changes may need to be made.

The following are required to get both the UPF and baseline model to compile.

• boost (recent, works with >= 1.56)
• Thrift >= 0.9.3
• gsl == 1.16
• cblas
• libarchive == 3.1.2
• redis >= 3.0.0
• hiredis == 0.13.3
• GoogleLOG (GLOG) == 0.3.3

As this project shares code with other projects, there are some optional dependencies.

• libLBFGS == 1.10
  • Github
• hihiredis
• atlas

There are a number of dependencies that are included in this repo directly. They are compiled on-demand.

• Google test (included in repo)
• Eigen (included in repo)
• Concrete >= 4.8 < 5
  • Concrete is a data schema, described in Ferraro et al., 2014.

Depending on where you installed the dependencies, you may need to update Makefile.config.

• If the headers are installed in /usr/local/include, and the shared objects are installed in /usr/local/lib, then you do not have to change anything.
• If the headers (and shared objects) are installed in the same place (but not /usr/local/{include,lib}), then change lines 5 and 6 of Makefile.config.
• If the headers (shared objects) are installed in separate directories, then change lines 11-61 as appropriate.

make help will display all known targets. To build the UPF model, run make models/upf_cgibbs_driver. To build the baseline model, run make models/crtlda_cgibbs_basic_driver.

There are a number of Makefile ENV variables that can be set to change compilation. Some major ones are:

• DEBUG={0,1}: turn on debug compilation. DEBUG=1 turns ON debugging, i.e., -g -O0). To run quickly, DEBUG=0. Default: 0
• LINK_HOW={dynamic,static}: use dynamic or static linking for certain libraries. Default: dynamic
• LOG_AS_COUT: This is meant to change what logging is used. It is by default undefined. Set it to anything to use stdout logging instead of Google Log.

Due to licensing issues, I cannot release the fully annotated input files. They are Concrete Communications with:

• a Stanford dependency parse (collapsed-cc)
• Stanford part-of-speech and lemmatization tags
• Stanford entity coreference
• Semafor frame semantic parses.

The list of training ids is in data/nyt10k.id_list.txt.

This code is released under GPL v3.0. Please contact me (ferraro [at] cs [dot] jhu [dot] edu) with any questions.