Merge remote-tracking branch 'upstream/master'

3 files changed, 175 insertions, 0 deletions

diff --git a/python/cdec/scfg/__init__.py b/python/cdec/scfg/__init__.py
new file mode 100644
index 00000000..6eb2f88f
--- /dev/null
+++ b/python/cdec/scfg/__init__.py
@@ -0,0 +1 @@
+from extractor import GrammarExtractor
diff --git a/python/cdec/scfg/extractor.py b/python/cdec/scfg/extractor.py
new file mode 100644
index 00000000..9f1e1137
--- /dev/null
+++ b/python/cdec/scfg/extractor.py
@@ -0,0 +1,112 @@
+#!/usr/bin/env python
+import StringIO
+
+import clex
+import rulefactory
+import calignment
+import csuf
+import cdat
+import sym
+import log
+
+log.level = -1
+
+from features import EgivenFCoherent, SampleCountF, CountEF,\
+        MaxLexEgivenF, MaxLexFgivenE, IsSingletonF, IsSingletonFE
+from features import contextless
+
+class Output(StringIO.StringIO):
+    def close(self):
+        pass
+
+    def __str__(self):
+        return self.getvalue()
+
+from itertools import chain
+
+def get_cn(sentence):
+    sentence = chain(('<s>',), sentence.split(), ('</s>',))
+    sentence = (sym.fromstring(word, terminal=True) for word in sentence)
+    return tuple(((word, None, 1), ) for word in sentence)
+
+class PhonyGrammar:
+    def add(self, thing):
+        pass
+
+class GrammarExtractor:
+    def __init__(self, config):
+        alignment = calignment.Alignment(config['a_file'], from_binary=True)
+        self.factory = rulefactory.HieroCachingRuleFactory(
+                # compiled alignment object (REQUIRED)
+                alignment=alignment,
+                # name of generic nonterminal used by Hiero
+                category="[X]",
+                # do not change for extraction
+                grammar=PhonyGrammar(), # TODO: set to None?
+                # maximum number of contiguous chunks of terminal symbols in RHS of a rule. If None, defaults to max_nonterminals+1
+                max_chunks=None,
+                # maximum span of a grammar rule in TEST DATA
+                max_initial_size=15,
+                # maximum number of symbols (both T and NT) allowed in a rule
+                max_length=config['max_len'],
+                # maximum number of nonterminals allowed in a rule (set >2 at your own risk)
+                max_nonterminals=config['max_nt'],
+                # maximum number of contiguous chunks of terminal symbols in target-side RHS of a rule. If None, defaults to max_nonterminals+1
+                max_target_chunks=None,
+                # maximum number of target side symbols (both T and NT) allowed in a rule. If None, defaults to max_initial_size
+                max_target_length=None,
+                # minimum span of a nonterminal in the RHS of a rule in TEST DATA
+                min_gap_size=1,
+                # filename of file containing precomputed collocations
+                precompute_file=config['precompute_file'],
+                # maximum frequency rank of patterns used to compute triples (don't set higher than 20).
+                precompute_secondary_rank=config['rank2'],
+                # maximum frequency rank of patterns used to compute collocations (no need to set higher than maybe 200-300)
+                precompute_rank=config['rank1'],
+                # require extracted rules to have at least one aligned word
+                require_aligned_terminal=True,
+                # require each contiguous chunk of extracted rules to have at least one aligned word
+                require_aligned_chunks=False,
+                # generate a complete grammar for each input sentence
+                per_sentence_grammar=True,
+                # maximum span of a grammar rule extracted from TRAINING DATA
+                train_max_initial_size=config['max_size'],
+                # minimum span of an RHS nonterminal in a rule extracted from TRAINING DATA
+                train_min_gap_size=config['min_gap'],
+                # True if phrases should be tight, False otherwise (False seems to give better results but is slower)
+                tight_phrases=True,
+                )
+        self.fsarray = csuf.SuffixArray(config['f_sa_file'], from_binary=True)
+        self.edarray = cdat.DataArray(config['e_file'], from_binary=True)
+
+        self.factory.registerContext(self)
+
+        # lower=faster, higher=better; improvements level off above 200-300 range, -1 = don't sample, use all data (VERY SLOW!)
+        self.sampler = rulefactory.Sampler(300)
+        self.sampler.registerContext(self)
+
+        # lexical weighting tables
+        tt = clex.CLex(config['lex_file'], from_binary=True)
+
+        self.models = (EgivenFCoherent, SampleCountF, CountEF, 
+                MaxLexFgivenE(tt), MaxLexEgivenF(tt), IsSingletonF, IsSingletonFE)
+        self.models = tuple(contextless(feature) for feature in self.models)
+
+    def grammar(self, sentence):
+        out = Output()
+        cn = get_cn(sentence)
+        self.factory.input_file(cn, out)
+        return str(out)
+
+def main(config):
+    sys.path.append(os.path.dirname(config))
+    module =  __import__(os.path.basename(config).replace('.py', ''))
+    extractor = GrammarExtractor(module.__dict__)
+    print extractor.grammar(next(sys.stdin))
+
+if __name__ == '__main__':
+    import sys, os
+    if len(sys.argv) != 2 or not sys.argv[1].endswith('.py'):
+        sys.stderr.write('Usage: %s config.py\n' % sys.argv[0])
+        sys.exit(1)
+    main(*sys.argv[1:])
diff --git a/python/cdec/scfg/features.py b/python/cdec/scfg/features.py
new file mode 100644
index 00000000..6419cdd8
--- /dev/null
+++ b/python/cdec/scfg/features.py
@@ -0,0 +1,62 @@
+from __future__ import division
+import math
+import sym
+
+def contextless(feature):
+    feature.compute_contextless_score = feature
+    return feature
+
+MAXSCORE = 99
+
+def EgivenF(fphrase, ephrase, paircount, fcount, fsample_count): # p(e|f)
+    return -math.log10(paircount/fcount)
+
+def CountEF(fphrase, ephrase, paircount, fcount, fsample_count):
+    return math.log10(1 + paircount)
+
+def SampleCountF(fphrase, ephrase, paircount, fcount, fsample_count):
+    return math.log10(1 + fsample_count)
+
+def EgivenFCoherent(fphrase, ephrase, paircount, fcount, fsample_count):
+    prob = paircount/fsample_count
+    return -math.log10(prob) if prob > 0 else MAXSCORE
+
+def CoherenceProb(fphrase, ephrase, paircount, fcount, fsample_count):
+    return -math.log10(fcount/fsample_count)
+
+def MaxLexEgivenF(ttable):
+    def feature(fphrase, ephrase, paircount, fcount, fsample_count):
+        fwords = [sym.tostring(w) for w in fphrase if not sym.isvar(w)] + ['NULL']
+        ewords = (sym.tostring(w) for w in ephrase if not sym.isvar(w))
+        def score():
+            for e in ewords:
+              maxScore = max(ttable.get_score(f, e, 0) for f in fwords)
+              yield -math.log10(maxScore) if maxScore > 0 else MAXSCORE
+        return sum(score())
+    return feature
+
+def MaxLexFgivenE(ttable):
+    def feature(fphrase, ephrase, paircount, fcount, fsample_count):
+        fwords = (sym.tostring(w) for w in fphrase if not sym.isvar(w))
+        ewords = [sym.tostring(w) for w in ephrase if not sym.isvar(w)] + ['NULL']
+        def score():
+            for f in fwords:
+              maxScore = max(ttable.get_score(f, e, 1) for e in ewords)
+              yield -math.log10(maxScore) if maxScore > 0 else MAXSCORE
+        return sum(score())
+    return feature
+
+def IsSingletonF(fphrase, ephrase, paircount, fcount, fsample_count):
+    return (fcount == 1)
+
+def IsSingletonFE(fphrase, ephrase, paircount, fcount, fsample_count):
+    return (paircount == 1)
+
+def IsNotSingletonF(fphrase, ephrase, paircount, fcount, fsample_count):
+    return (fcount > 1)
+
+def IsNotSingletonFE(fphrase, ephrase, paircount, fcount, fsample_count):
+    return (paircount > 1)
+
+def IsFEGreaterThanZero(fphrase, ephrase, paircount, fcount, fsample_count):
+    return (paircount > 0.01)