From 98e13188c40adf13b73ca994d3410f3e87d20355 Mon Sep 17 00:00:00 2001 From: Victor Chahuneau Date: Mon, 4 Jun 2012 02:16:59 -0400 Subject: Python module: initial version --- python/cdec/scfg/__init__.py | 1 + python/cdec/scfg/extractor.py | 112 ++++++++++++++++++++++++++++++++++++++++++ python/cdec/scfg/features.py | 62 +++++++++++++++++++++++ 3 files changed, 175 insertions(+) create mode 100644 python/cdec/scfg/__init__.py create mode 100644 python/cdec/scfg/extractor.py create mode 100644 python/cdec/scfg/features.py (limited to 'python/cdec/scfg') 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(('',), sentence.split(), ('',)) + 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) -- cgit v1.2.3