lopez suffix array extractor with copyrighted david chiang code excised

1 files changed, 116 insertions, 0 deletions

diff --git a/sa-extract/extract.ini b/sa-extract/extract.ini
new file mode 100644
index 00000000..56913245
--- /dev/null
+++ b/sa-extract/extract.ini
@@ -0,0 +1,116 @@
+# This .ini file extracts grammars to a file using
+# the pattern matching infrastructure.
+#
+# Does not do any decoding.
+#
+# Variables can be set using sa-system.pl
+#
+# Usage: decoder.py -c <this ini file> [-x <grammar file>]
+#
+# If the -x option is used, grammar will be written to the
+# specified file, otherwise it is written to $PWD/grammar.out
+# 
+# NOTE: all information about rules is cached, so use generous
+# memory limits (rules themselves are not cached.)
+
+import os
+import manager
+import clex
+import context_model
+import rulefactory
+import calignment
+import sys
+
+out_grammar_file = "grammar.out"
+if opts.extra:
+	out_grammar_file = opts.extra
+
+# *** these variables written by sa-system.pl.  Do not modify ***
+lm_file = "/tmp/sa-redpony/de-en/lm/lm/lm.gz"
+f_sa_file = "/tmp/sa-redpony/de-en/bitext/soseos.vc/f.sa.bin"
+e_file = "/tmp/sa-redpony/de-en/bitext/soseos.vc/e.bin"
+a_file = "/tmp/sa-redpony/de-en/bitext/soseos.vc/a/gdfa/a.bin"
+lex_file = "/tmp/sa-redpony/de-en/bitext/soseos.vc/a/gdfa/lex.bin"
+max_len = 5
+max_nt = 2
+max_size=10
+min_gap=1
+rank1 = 100
+rank2 = 10
+precompute_file = "/tmp/sa-redpony/de-en/bitext/soseos.vc/precomp.5.2.10.1.100.10.bin"
+
+# check for path errors
+if not os.path.exists(f_sa_file):
+	raise Exception("Cannot find compiled source language suffix array file %s" % f_sa_file)
+if not os.path.exists(e_file):
+	raise Exception("Cannot find compiled target language array file %s" % e_file)
+if not os.path.exists(a_file):
+	raise Exception("Cannot find compiled alignment file %s" % a_file)
+if not os.path.exists(lex_file):
+	raise Exception("Cannot find compiled lexical weights file %s" % lex_file)
+if not os.path.exists(precompute_file):
+	log.writeln("Could not find precomputed collocations %s, decoding will be slower" % precompute_file)
+	precompute_file = None
+
+### Output options
+mark_phrases = False      # show derivation as SGML markup in output
+mert_mark_phrases = False # do the same when generating n-best lists (don't use this with minimum error rate training!)
+
+# Verbosity. 0 = silent, 1 = normal, 2-5 = verbose
+log.level = 1
+log.file = sys.stderr
+
+# pattern-matching stuff
+class PhonyGrammar:  # saves us the cost of keeping the rules around
+	def add(self, thing):
+		pass
+
+local_grammar = PhonyGrammar()
+xcat="X"
+
+cm = manager.ContextManager(
+	f_sa_file,
+	e_file,
+	sampler=rulefactory.Sampler(300),  # lower=faster, higher=better; improvements level off above 200-300 range, -1 = don't sample, use all data (VERY SLOW!)
+	rulefactory=rulefactory.HieroCachingRuleFactory(
+		alignment=calignment.Alignment(  # compiled alignment object (REQUIRED)
+			a_file, 
+			from_binary=True
+			),
+		category="["+xcat+"]",           # name of generic nonterminal used by Hiero
+		grammar=local_grammar,           # do not change for extraction
+		max_chunks=None,                 # maximum number of contiguous chunks of terminal symbols in RHS of a rule. If None, defaults to max_nonterminals+1
+		max_initial_size=15,             # maximum span of a grammar rule in TEST DATA
+		max_length=max_len,              # maximum number of symbols (both T and NT) allowed in a rule
+		max_nonterminals=max_nt,         # maximum number of nonterminals allowed in a rule (set >2 at your own risk)
+		max_target_chunks=None,          # maximum number of contiguous chunks of terminal symbols in target-side RHS of a rule. If None, defaults to max_nonterminals+1
+		max_target_length=None,          # maximum number of target side symbols (both T and NT) allowed in a rule. If None, defaults to max_initial_size
+		min_gap_size=1,                  # minimum span of a nonterminal in the RHS of a rule in TEST DATA
+		precompute_file=precompute_file, # filename of file containing precomputed collocations
+		precompute_secondary_rank=rank2, # maximum frequency rank of patterns used to compute triples (don't set higher than 20).
+		precompute_rank=rank1,           # maximum frequency rank of patterns used to compute collocations (no need to set higher than maybe 200-300)
+		require_aligned_terminal=True,   # require extracted rules to have at least one aligned word
+		require_aligned_chunks=False,    # require each contiguous chunk of extracted rules to have at least one aligned word
+		per_sentence_grammar=True,       # generate a complete grammar for each input sentence
+		rule_file=out_grammar_file,      # grammar is written to this file (sentence id is added to file name for per sentence grammars)
+		train_max_initial_size=max_size, # maximum span of a grammar rule extracted from TRAINING DATA
+		train_min_gap_size=min_gap,      # minimum span of an RHS nonterminal in a rule extracted from TRAINING DATA
+		tight_phrases=True,              # True if phrases should be tight, False otherwise (False seems to give better results but is slower)
+		),
+	from_binary=True
+	)
+
+# lexical weighting tables
+tt = clex.CLex(lex_file, from_binary=True)
+
+# Only include features that depend on rule identity here
+add_model(context_model.EgivenFCoherent(cm), 0.125) 
+add_model(context_model.SampleCountF(cm), 0.125)
+add_model(context_model.CountEF(cm), 0.125)
+add_model(context_model.MaxLexFgivenE(cm, tt), 0.125) 
+add_model(context_model.MaxLexEgivenF(cm, tt), 0.125) 
+add_model(context_model.IsSingletonF(cm), 0.125)
+add_model(context_model.IsSingletonFE(cm), 0.125)
+
+# grammars, search parameters and all that other stuff are irrelevant
+