1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
|
#!/usr/bin/env python
# Author: Phil Williams
# Usage: filter-rule-table.py [--min-non-initial-rule-count=N] INPUT
#
# Given a rule table (on stdin) and an input text, filter out rules that
# couldn't be used in parsing the input and write the resulting rule table
# to stdout. The input text is assumed to contain the same factors as
# the rule table and is assumed to be small (not more than a few thousand
# sentences): the current algorithm won't scale well to large input sets.
#
# The filtering algorithm considers a source RHS to be a sequence of
# words and gaps, which must match a sequence of words in one of the
# input sentences, with at least one input word per gap. The NT labels
# are ignored, so for example a rule with the source RHS "the JJ dog"
# would be allowed if the sequence "the slobbering dog" occurs in one of
# the input sentences, even if there's no rule to derive a JJ from
# "slobbering." (If "slobbering" were an unknown word, the 'unknown-lhs'
# decoder option would allow it to take a number of NT labels, likely
# including JJ, with varying probabilities, so removing the rule would
# be a bad idea.)
import optparse
import sys
class NGram(tuple):
pass
class Gap:
def __init__(self, minSpan):
self.minSpan = minSpan
def getMinSpan(self):
return self.minSpan
def printUsage():
sys.stderr.write("Usage: filter-rule-table.py [--min-non-initial-rule-count=N] INPUT")
def main():
parser = optparse.OptionParser()
parser.add_option("-c", "--min-non-initial-rule-count",
action="store", dest="minCount",
type="float", default="0.0",
help="prune non-initial rules where count is below N",
metavar="N")
(options, args) = parser.parse_args()
if len(args) != 1:
printUsage()
sys.exit(1)
N = 7
inputSentences = []
for line in open(args[0]):
inputSentences.append(line.split())
filterRuleTable(sys.stdin, inputSentences, N, options)
def filterRuleTable(ruleTable, inputSentences, N, options):
# Map each input n-gram (n = 1..N) to a map from sentence indices to
# lists of intra-sentence indices.
occurrences = {}
for i, sentence in enumerate(inputSentences):
for n in range(1, N+1):
for j in range(0, len(sentence)-n+1):
ngram = NGram(sentence[j:j+n])
innerMap = occurrences.setdefault(ngram, {})
indices = innerMap.setdefault(i, [])
indices.append(j)
# Compare rules against input.
prevRHS = None
prevRuleIncluded = None
for line in ruleTable:
rhs, count = parseRule(line)
# Prune non-initial rule if count is below threshold.
if count != None and count < options.minCount and isNonInitialRule(rhs):
if prevRHS != rhs:
prevRuleIncluded = None
prevRHS = rhs
continue
# If source RHS is same as last rule's then we already know whether to
# filter or not (unless it was pruned before checking).
if rhs == prevRHS and prevRuleIncluded != None:
if prevRuleIncluded:
print line,
continue
prevRHS = rhs
# <s> and </s> can appear in glue rules.
if rhs[0] == "<s>" or rhs[-1] == "</s>":
print line,
prevRuleIncluded = True
continue
segments = segmentRHS(rhs, N)
ngramMaps = [occurrences.get(s, {}) for s in segments if isinstance(s, NGram)]
if len(ngramMaps) == 0:
print line,
prevRuleIncluded = True
continue
# Determine the sentences in which all n-grams co-occur.
sentences = set(ngramMaps[0].keys())
for ngramMap in ngramMaps[1:]:
sentences &= set(ngramMap.keys())
# Try to match rule in candidate sentences.
match = False
for sentenceIndex in sentences:
sentenceLength = len(inputSentences[sentenceIndex])
for indexSeq in enumerateIndexSeqs(ngramMaps, sentenceIndex, 0):
if matchSegments(segments, indexSeq, sentenceLength):
print line,
match = True
break
if match:
break
prevRuleIncluded = match
# Parse a line of the rule table and return a tuple containing two items,
# the list of RHS source symbols and the rule count (if present).
def parseRule(line):
cols = line.split("|||")
rhsSourceSymbols = cols[0].split()[:-1]
ruleCount = None
if len(cols) > 4:
counts = cols[4].split()
if len(counts) == 3:
ruleCount = float(counts[2])
return (rhsSourceSymbols, ruleCount)
def isNT(symbol):
return symbol[0] == '[' and symbol[-1] == ']'
def isNonInitialRule(rhs):
for symbol in rhs:
if isNT(symbol):
return True
return False
def segmentRHS(rhs, N):
segments = []
terminals = []
minGapWidth = 0
for symbol in rhs:
if isNT(symbol):
if len(terminals) > 0:
assert minGapWidth == 0
segments.append(NGram(terminals))
terminals = []
minGapWidth += 1
else:
if minGapWidth > 0:
assert len(terminals) == 0
segments.append(Gap(minGapWidth))
minGapWidth = 0
terminals.append(symbol)
if len(terminals) == N:
segments.append(NGram(terminals))
terminals = []
if minGapWidth > 0:
assert len(terminals) == 0
segments.append(Gap(minGapWidth))
elif len(terminals) > 0:
segments.append(NGram(terminals))
return segments
def matchSegments(segments, indexSeq, sentenceLength):
assert len(segments) > 0
firstSegment = segments[0]
i = 0
if isinstance(firstSegment, Gap):
minPos = firstSegment.getMinSpan()
maxPos = sentenceLength-1
else:
minPos = indexSeq[i] + len(firstSegment)
i += 1
maxPos = minPos
for segment in segments[1:]:
if isinstance(segment, Gap):
if minPos + segment.getMinSpan() > sentenceLength:
return False
minPos = minPos + segment.getMinSpan()
maxPos = sentenceLength-1
else:
pos = indexSeq[i]
i += 1
if pos < minPos or pos > maxPos:
return False
minPos = pos + len(segment)
maxPos = minPos
return True
def enumerateIndexSeqs(ngramMaps, sentenceIndex, minFirstIndex):
assert len(ngramMaps) > 0
if len(ngramMaps) == 1:
for index in ngramMaps[0][sentenceIndex]:
if index >= minFirstIndex:
yield [index]
return
for index in ngramMaps[0][sentenceIndex]:
if index < minFirstIndex:
continue
for seq in enumerateIndexSeqs(ngramMaps[1:], sentenceIndex, index+1):
assert seq[0] > index
yield [index] + seq
if __name__ == "__main__":
main()
|