# MERT optimized configuration
# decoder /export/projects/mpost/language-packs/es-en/5/tune/model/run-joshua.sh
# BLEU 0.2655 on dev /export/projects/mpost/language-packs/es-en/5/data/tune/corpus.es
# We were before running iteration 4
# finished Tue Sep 27 12:03:42 EDT 2016
# This file is a template for the Joshua pipeline; variables enclosed
# in <angle-brackets> are substituted by the pipeline script as
# appropriate.  This file also serves to document Joshua's many
# parameters.

# These are the grammar file specifications.  Joshua supports an
# arbitrary number of grammar files, each specified on its own line
# using the following format:
#
#   tm = TYPE OWNER LIMIT FILE
# 
# TYPE is "packed", "thrax", or "samt".  The latter denotes the format
# used in Zollmann and Venugopal's SAMT decoder
# (http://www.cs.cmu.edu/~zollmann/samt/).
# 
# OWNER is the "owner" of the rules in the grammar; this is used to
# determine which set of phrasal features apply to the grammar's
# rules.  Having different owners allows different features to be
# applied to different grammars, and for grammars to share features
# across files.
#
# LIMIT is the maximum input span permitted for the application of
# grammar rules found in the grammar file.  A value of -1 implies no limit.
#
# FILE is the grammar file (or directory when using packed grammars).
# The file can be compressed with gzip, which is determined by the
# presence or absence of a ".gz" file extension.
#
# By a convention defined by Chiang (2007), the grammars are split
# into two files: the main translation grammar containing all the
# learned translation rules, and a glue grammar which supports
# monotonic concatenation of hierarchical phrases. The glue grammar's
# main distinction from the regular grammar is that the span limit
# does not apply to it.  

tm = phrase -owner pt -maxspan 0 -path model/grammar.gz.packed

# This symbol is used over unknown words in the source language

default-non-terminal = X

# This is the goal nonterminal, used to determine when a complete
# parse is found.  It should correspond to the root-level rules in the
# glue grammar.

goal-symbol = GOAL

# Language model config.
#
# Multiple language models are supported.  For each language model,
# create one of the following lines:
#
# feature-function = LanguageModel -lm_type TYPE -lm_order ORDER -lm_file FILE
# feature-function = StateMinimizingLanguageModel -lm_order ORDER -lm_file FILE
#
# - TYPE is one of "kenlm" or "berkeleylm"
# - ORDER is the order of the language model (default 5)
# - FILE is the path to the LM file. This can be binarized if appropriate to the type
#   (e.g., KenLM has a compiled format)
#
# A state-minimizing LM collapses left-state. Currently only KenLM supports this.
#
# For each LM, add a weight lm_INDEX below, where indexing starts from 0.



# The suffix _OOV is appended to unknown source-language words if this
# is set to true.

mark-oovs = false

# The search algorithm: "cky" for hierarchical / phrase-based decoding, 
# "stack" for phrase-based decoding
search = stack

# The pop-limit for decoding.  This determines how many hypotheses are
# considered over each span of the input.

pop-limit = 100

# How many hypotheses to output

top-n = 1

# Whether those hypotheses should be distinct strings

use-unique-nbest = true

# This is the default format of the ouput printed to STDOUT.  The variables that can be
# substituted are:
#
# %i: the sentence number (0-indexed)
# %s: the translated sentence
# %t: the derivation tree
# %f: the feature string
# %c: the model cost

output-format = %S

# When printing the trees (%t in 'output-format'), this controls whether the alignments
# are also printed.

include-align-index = false

# And these are the feature functions to activate.
feature-function = OOVPenalty
feature-function = WordPenalty

## Model weights #####################################################

# For each langage model line listed above, create a weight in the
# following format: the keyword "lm", a 0-based index, and the weight.
# lm_INDEX WEIGHT


# The phrasal weights correspond to weights stored with each of the
# grammar rules.  The format is
#
#   tm_OWNER_COLUMN WEIGHT
#
# where COLUMN denotes the 0-based order of the parameter in the
# grammar file and WEIGHT is the corresponding weight.  In the future,
# we plan to add a sparse feature representation which will simplify
# this.

# The wordpenalty feature counts the number of words in each hypothesis.


# This feature counts the number of unknown words in the hypothesis.


# This feature weights paths through an input lattice.  It is only activated
# when decoding lattices.
feature-function = LanguageModel -lm_order 4 -lm_file model/lm.gz -lm_type berkeleylm
feature-function = Distortion
feature-function = PhrasePenalty



lowercase = -project-case

lm_0 0.242132004556722
WordPenalty -0.111308832033767
OOVPenalty 0.0101534888932218
tm_pt_2 0.0241130425384253
PhrasePenalty -0.0240605834315291
tm_pt_0 0.0262269656358665
tm_pt_1 0.0535319307753204
Distortion 0.121100027216756
tm_pt_3 0.191275104853519
tm_pt_4 0.119489193983075
tm_pt_5 0.076608826081799