summaryrefslogtreecommitdiff
path: root/python/src/sa/online_extractor.py
blob: 0c013cb35f04228c2fc5e9e859ee5fda3595d040 (plain)
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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
#!/usr/bin/env python

import collections, sys

import cdec.configobj

CAT = '[X]'  # Default non-terminal
MAX_SIZE = 15  # Max span of a grammar rule (source)
MAX_LEN = 5  # Max number of terminals and non-terminals in a rule (source)
MAX_NT = 2  # Max number of non-terminals in a rule
MIN_GAP = 1  # Min number of terminals between non-terminals (source)

# Spans are _inclusive_ on both ends [i, j]
# TODO: Replace all of this with bit vectors?
def span_check(vec, i, j):
    k = i
    while k <= j:
        if vec[k]:
            return False
        k += 1
    return True

def span_flip(vec, i, j):
    k = i
    while k <= j:
        vec[k] = ~vec[k]
        k += 1

# Next non-terminal
def next_nt(nt):
     if not nt:
         return 1
     return nt[-1][0] + 1

class NonTerminal:
    def __init__(self, index):
        self.index = index
    def __str__(self):
        return '[X,{0}]'.format(self.index)

def fmt_rule(f_sym, e_sym, links):
    a_str = ' '.join('{0}-{1}'.format(i, j) for (i, j) in links)
    return '[X] ||| {0} ||| {1} ||| {2}'.format(' '.join(str(sym) for sym in f_sym),
                                                ' '.join(str(sym) for sym in e_sym),
                                                a_str)

# Create a rule from source, target, non-terminals, and alignments
def form_rules(f_i, e_i, f_span, e_span, nt, al):

    # This could be more efficient but is unlikely to be the bottleneck

    rules = []

    nt_inv = sorted(nt, cmp=lambda x, y: cmp(x[3], y[3]))

    f_sym = f_span[:]
    off = f_i
    for next_nt in nt:
        nt_len = (next_nt[2] - next_nt[1]) + 1
        i = 0
        while i < nt_len:
            f_sym.pop(next_nt[1] - off)
            i += 1
        f_sym.insert(next_nt[1] - off, NonTerminal(next_nt[0]))
        off += (nt_len - 1)

    e_sym = e_span[:]
    off = e_i
    for next_nt in nt_inv:
        nt_len = (next_nt[4] - next_nt[3]) + 1
        i = 0
        while i < nt_len:
            e_sym.pop(next_nt[3] - off)
            i += 1
        e_sym.insert(next_nt[3] - off, NonTerminal(next_nt[0]))
        off += (nt_len - 1)

    # Adjusting alignment links takes some doing
    links = [list(link) for sub in al for link in sub]
    links_len = len(links)
    nt_len = len(nt)
    nt_i = 0
    off = f_i
    i = 0
    while i < links_len:
        while nt_i < nt_len and links[i][0] > nt[nt_i][1]:
            off += (nt[nt_i][2] - nt[nt_i][1])
            nt_i += 1
        links[i][0] -= off
        i += 1
    nt_i = 0
    off = e_i
    i = 0
    while i < links_len:
        while nt_i < nt_len and links[i][1] > nt_inv[nt_i][3]:
            off += (nt_inv[nt_i][4] - nt_inv[nt_i][3])
            nt_i += 1
        links[i][1] -= off
        i += 1

    rules.append(fmt_rule(f_sym, e_sym, links))
    return rules

class OnlineGrammarExtractor:

    def __init__(self, config=None):
        if isinstance(config, str) or isinstance(config, unicode):
            if not os.path.exists(config):
                raise IOError('cannot read configuration from {0}'.format(config))
            config = cdec.configobj.ConfigObj(config, unrepr=True)
        elif not config:
            config = collections.defaultdict(lambda: None)
        self.category = CAT
        self.max_size = MAX_SIZE
        self.max_length = config['max_len'] or MAX_LEN
        self.max_nonterminals = config['max_nt'] or MAX_NT
        self.min_gap_size = MIN_GAP
        # Hard coded: require at least one aligned word
        # Hard coded: require tight phrases

        # Phrase counts
        self.phrases_f = collections.defaultdict(lambda: 0)
        self.phrases_e = collections.defaultdict(lambda: 0)
        self.phrases_fe = collections.defaultdict(lambda: collections.defaultdict(lambda: 0))

        # Bilexical counts
        self.bilex_f = collections.defaultdict(lambda: 0)
        self.bilex_e = collections.defaultdict(lambda: 0)
        self.bilex_fe = collections.defaultdict(lambda: collections.defaultdict(lambda: 0))

    # Aggregate bilexical counts
    def aggr_bilex(self, f_words, e_words):

        for e_w in e_words:
            self.bilex_e[e_w] += 1

        for f_w in f_words:
            self.bilex_f[f_w] += 1
            for e_w in e_words:
                self.bilex_fe[f_w][e_w] += 1

    # Aggregate stats from a training instance:
    # Extract hierarchical phrase pairs
    # Update bilexical counts
    def add_instance(self, f_words, e_words, alignment):

        # Bilexical counts
        self.aggr_bilex(f_words, e_words)

        # Phrase pairs extracted from this instance
        phrases = set()

        f_len = len(f_words)

        # Pre-compute alignment info
        al = [[] for i in range(f_len)]
        al_span = [[f_len + 1, -1] for i in range(f_len)]
        for (f, e) in alignment:
            al[f].append(e)
            al_span[f][0] = min(al_span[f][0], e)
            al_span[f][1] = max(al_span[f][1], e)

        # Target side word coverage
        # TODO: Does Cython do bit vectors?
        cover = [0] * f_len

        # Extract all possible hierarchical phrases starting at a source index
        # f_ i and j are current, e_ i and j are previous
        def extract(f_i, f_j, e_i, e_j, wc, links, nt, nt_open):
            # Phrase extraction limits
            if wc > self.max_length or (f_j + 1) >= f_len or \
                    (f_j - f_i) + 1 > self.max_size:
                return
            # Unaligned word
            if not al[f_j]:
                # Open non-terminal: extend
                if nt_open:
                    nt[-1][2] += 1
                    extract(f_i, f_j + 1, e_i, e_j, wc, links, nt, True)
                    nt[-1][2] -= 1
                # No open non-terminal: extend with word
                else:
                    extract(f_i, f_j + 1, e_i, e_j, wc + 1, links, nt, False)
                return
            # Aligned word
            link_i = al_span[f_j][0]
            link_j = al_span[f_j][1]
            new_e_i = min(link_i, e_i)
            new_e_j = max(link_j, e_j)
            # Open non-terminal: close, extract, extend
            if nt_open:
                # Close non-terminal, checking for collisions
                old_last_nt = nt[-1][:]
                nt[-1][2] = f_j
                if link_i < nt[-1][3]:
                    if not span_check(cover, link_i, nt[-1][3] - 1):
                        nt[-1] = old_last_nt
                        return
                    span_flip(cover, link_i, nt[-1][3] - 1)
                    nt[-1][3] = link_i
                if link_j > nt[-1][4]:
                    if not span_check(cover, nt[-1][4] + 1, link_j):
                        nt[-1] = old_last_nt
                        return
                    span_flip(cover, nt[-1][4] + 1, link_j)
                    nt[-1][4] = link_j
                for rule in form_rules(f_i, new_e_i, f_words[f_i:f_j + 1], e_words[new_e_i:new_e_j + 1], nt, links):
                    phrases.add(rule)
                extract(f_i, f_j + 1, new_e_i, new_e_j, wc, links, nt, False)
                nt[-1] = old_last_nt
                if link_i < nt[-1][3]:
                    span_flip(cover, link_i, nt[-1][3] - 1)
                if link_j > nt[-1][4]:
                    span_flip(cover, nt[-1][4] + 1, link_j)
                return
            # No open non-terminal
            # Extract, extend with word
            collision = False
            for link in al[f_j]:
                if cover[link]:
                    collision = True
            # Collisions block extraction and extension, but may be okay for
            # continuing non-terminals
            if not collision:
                plus_links = []
                for link in al[f_j]:
                    plus_links.append((f_j, link))
                    cover[link] = ~cover[link]
                links.append(plus_links)
                for rule in form_rules(f_i, new_e_i, f_words[f_i:f_j + 1], e_words[new_e_i:new_e_j + 1], nt, links):
                    phrases.add(rule)
                extract(f_i, f_j + 1, new_e_i, new_e_j, wc + 1, links, nt, False)
                links.pop()
                for link in al[f_j]:
                    cover[link] = ~cover[link]
            # Try to add a word to a (closed) non-terminal, extract, extend
            if nt and nt[-1][2] == f_j - 1:
                # Add to non-terminal, checking for collisions
                old_last_nt = nt[-1][:]
                nt[-1][2] = f_j
                if link_i < nt[-1][3]:
                    if not span_check(cover, link_i, nt[-1][3] - 1):
                        nt[-1] = old_last_nt
                        return
                    span_flip(cover, link_i, nt[-1][3] - 1)
                    nt[-1][3] = link_i
                if link_j > nt[-1][4]:
                    if not span_check(cover, nt[-1][4] + 1, link_j):
                        nt[-1] = old_last_nt
                        return
                    span_flip(cover, nt[-1][4] + 1, link_j)
                    nt[-1][4] = link_j
                # Require at least one word in phrase
                if links:
                    for rule in form_rules(f_i, new_e_i, f_words[f_i:f_j + 1], e_words[new_e_i:new_e_j + 1], nt, links):
                        phrases.add(rule)
                extract(f_i, f_j + 1, new_e_i, new_e_j, wc, links, nt, False)
                nt[-1] = old_last_nt
                if new_e_i < nt[-1][3]:
                    span_flip(cover, link_i, nt[-1][3] - 1)
                if link_j > nt[-1][4]:
                    span_flip(cover, nt[-1][4] + 1, link_j)
            # Try to start a new non-terminal, extract, extend
            if (not nt or f_j - nt[-1][2] > 1) and len(nt) < self.max_nonterminals:
                # Check for collisions
                if not span_check(cover, link_i, link_j):
                    return
                span_flip(cover, link_i, link_j)
                nt.append([next_nt(nt), f_j, f_j, link_i, link_j])
                # Require at least one word in phrase
                if links:
                    for rule in form_rules(f_i, new_e_i, f_words[f_i:f_j + 1], e_words[new_e_i:new_e_j + 1], nt, links):
                        phrases.add(rule)
                extract(f_i, f_j + 1, new_e_i, new_e_j, wc, links, nt, False)
                nt.pop()
                span_flip(cover, link_i, link_j)
            # TODO: try adding NT to start, end, both
            # check: one aligned word on boundary that is not part of a NT

        # Try to extract phrases from every f index
        f_i = 0
        while f_i < f_len:
            # Skip if phrases won't be tight on left side
            if not al[f_i]:
                f_i += 1
                continue
            extract(f_i, f_i, f_len + 1, -1, 1, [], [], False)
            f_i += 1

        for rule in sorted(phrases):
            print rule

def main(argv):

    extractor = OnlineGrammarExtractor()

    for line in sys.stdin:
        f_words, e_words, a_str = (x.split() for x in line.split('|||'))
        alignment = sorted(tuple(int(y) for y in x.split('-')) for x in a_str)
        extractor.add_instance(f_words, e_words, alignment)

if __name__ == '__main__':
    main(sys.argv)