6.0 (December 2014)
===================

Joshua 6.0 introduces many extensions and improvements. They include:

- Phrase-based decoder

  Joshua now includes an unlexicalized phrase-based decoder. It reads
  Moses phrase tables directly. Support for training phrase-based
  models is included in the pipeline (via callouts to Moses)

- Significant speed improvements

  Joshua is a lot faster (TODO: quantify). The phrase-based decoder is
  on roughly par with Moses in terms of decoding speed: slower for
  small beam sizes and faster for larger (above 200). GHKM decoding is
  significantly faster as a result of the dot chart removal due to
  Sennrich (2014).

- Numerous bugfixes and other improvements

  Dynamic lattice-based OOV segmenting, improved progress meters and
  verbose output handling, support for alignments in phrase table,
  latest KenLM (including lmplz)
	
5.0 (August 16, 2013)
===================

The main features of this release are described in

  "Joshua 5.0: Sparser, Better, Faster, Server"
  Matt Post, Juri Ganitkevitch, Luke Orland, Jonny Weese, Yuan Cao, and Chris Callison-Burch. 
  ACL Workshop on Statistical Machine Translation. August, 2013.
  www.statmt.org/wmt13/pdf/WMT26.pdf

- Sparse feature implementation

  Joshua now uses sparse features natively, with support for (hundreds of) thousands of
  features and large-scale discriminative tuning via PRO (included) or kbMIRA (via Moses).

- Significant performance improvements

  Joshua is up to 6 times faster than the previous release. It scales to many threads
  easily, and the packed grammar and amortized sorting virtually remove model loading times.

- Added left-state LM minimization support (via KenLM)

  Tests show that Joshua has parity with Moses in terms of speed and search.

- Thrax 2.0

  Thrax 2.0 is significantly faster, uses less disk space, and has been tested on corpora
  of one hundred million sentence pairs.

- Server

  Joshua now includes a multithreaded TCP/IP server with round-robin scheduling among
  connections.

- Many, may bugfixes


4.0 (July 2, 2012) 
==================

The main features of this release are described in 

  "Joshua 4.0: Packing, PRO, and Paraphrasing."
  Juri Ganitkevitch, Yuan Cao, Jonny Weese, Matt Post, and Chris Callison-Burch. 
  NAACL Workshop on Statistical Machine Translation, June, 2012.

They include:

- Significantly improved and expanded documentation (both user and developer)

  See http://joshua-decoder.org/4.0 or ./joshua-decoder.org/4.0/index.html (local mirror)

- Synchronous parsing

  Joshua will compute the best synchronous derivation over a pair of
  sentences.  Pass the sentences in in the form 

    source sentence ||| target sentence

  and set the parameter "parse = true" (either from the config file or
  command-line).

- PRO implementation

  We include an implementation of Pairwise Ranking Optimization (PRO,
  Hopkins & May, EMNLP 2011).  It can be activated by passing "--tuner
  pro" to the pipeline script.

- Grammar packing

  We include an efficient grammar representation that can be used to
  greatly reduce the memory footprint of large grammars.

- Numerous bugfixes

== 3.2 (February 17, 2012) ======================================

- Pop-limit pruning.

  Pruning can now be specified with a single parameter "pop-limit"
  parameter, which limits the number of pops from the cube pruning
  candidate list at the span level.  This replaces the beam and
  threshold pruning that was governed by four parameters (fuzz1,
  fuzz2, relative_threshold, and max_n_rules), whose performance and
  interaction was somewhat difficult to characterize.  The pop-limit
  allows a simple relationship between decoding time and model score
  to be defined.

  Setting "pop-limit" in the configuration file or from the command
  line turns off beam-and-threshold pruning, and its use is
  recommended.  The default setting is to use a pop-limit of 100.

- Multiple language model support

  You can now specify an arbitrary number of language models.  See the
  documentation in

    $JOSHUA/scripts/training/templates/mert/joshua.config 

  for information on how to do this.  You can also specify multiple
  --lmfile flags to the pipeline.pl script.

- Multiple optimizer + test runs (--optimizer-runs N), averaging the
  results at the end (Clark et al., ACL 2011)

- Added support for BerkeleyLM (Pauls and Klein, ACL 2011)

- Support for lattice decoding (thanks to Lane Schwartz and the
  miniSCALE 2012 team)

- Pipeline script:

  - Removed all external dependencies (e.g., Moses, SRILM)

  - Reorganized the training data

  - Permit multiple test runs with subsequent --test FILE --name NAME
    calls to the pipeline

  - GIZA++ runs are parallelized if more than one thread is permitted
    (--threads N, N >=2 )

  - Numerous bugfixes

  - Hadoop cluster rollout is now a single instance (slower but
    doesn't require error-prone server setup)

- Parameters

  - Joshua now dies if it encounters unknown parameters on the command
    line or config file

  - Parameters are now normalized to remove hyphens (-) and
    underscores (_) and to flatten case, permitting you to specify any
    of, for example, {pop-limit, popLimit, pop_limit, ...}

- Lots of reorganization and purging of old code


3.1 
=============================

- Fixed multithreading.  Use -threads N from the command line or
  configuration file to spawn N parallel decoding threads.

- Configuration file parameters can now be overridden from the command
  line.  The format is

  -parameter value

  Among these must be the configuration file itself, which can be
  referred to with -config, or -c for short.

3.0
===

- Added the highly parameterizable Hadoop-based Thrax grammar
  extractor, which extracts both Hiero and SAMT grammars.

- Incorporated a black-box pipeline script at
  $JOSHUA/scripts/training/pipeline.pl

- Moved development to github.com.