Skip to content

Mulrel ranker

MulRelRanker(config, device)

Bases: torch.nn.Module

Source code in /home/docs/checkouts/readthedocs.org/user_builds/rel/envs/latest/lib/python3.7/site-packages/REL/mulrel_ranker.py 
 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
def __init__(self, config, device):
    super(MulRelRanker, self).__init__()
    self.config = config
    # self.embeddings = embeddings
    self.device = device
    self.max_dist = 1000
    self.ent_top_n = 1000
    self.ent_ent_comp = "bilinear"  # config.get('ent_ent_comp', 'bilinear')  # bilinear, trans_e, fbilinear

    self.att_mat_diag = torch.nn.Parameter(torch.ones(self.config["emb_dims"]))
    self.tok_score_mat_diag = torch.nn.Parameter(
        torch.ones(self.config["emb_dims"])
    )

    self.score_combine_linear_1 = torch.nn.Linear(2, self.config["hid_dims"])
    self.score_combine_act_1 = torch.nn.ReLU()
    self.score_combine_linear_2 = torch.nn.Linear(self.config["hid_dims"], 1)

    if self.config["use_local"]:
        self.ent_localctx_comp = torch.nn.Parameter(
            torch.ones(self.config["emb_dims"])
        )

    if self.config["use_pad_ent"]:
        self.pad_ent_emb = torch.nn.Parameter(
            torch.randn(1, self.config["emb_dims"]) * 0.1
        )
        self.pad_ctx_vec = torch.nn.Parameter(
            torch.randn(1, self.config["emb_dims"]) * 0.1
        )

    self.ctx_layer = torch.nn.Sequential(
        torch.nn.Linear(self.config["emb_dims"] * 3, self.config["emb_dims"]),
        torch.nn.Tanh(),
        torch.nn.Dropout(p=self.config["dropout_rate"]),
    )

    self.rel_embs = (
        torch.randn(self.config["n_rels"], self.config["emb_dims"]) * 0.01
    )
    self.rel_embs[0] = 1 + torch.randn(self.config["emb_dims"]) * 0.01
    self.rel_embs = torch.nn.Parameter(self.rel_embs)

    self.ew_embs = torch.nn.Parameter(
        torch.randn(self.config["n_rels"], self.config["emb_dims"]) * 0.01
    )
    self._coh_ctx_vecs = None

    self.score_combine = torch.nn.Sequential(
        torch.nn.Linear(2, self.config["hid_dims"]),
        torch.nn.ReLU(),
        torch.nn.Linear(self.config["hid_dims"], 1),
    )

__local_ent_scores(token_ids, tok_mask, entity_ids, entity_mask, embeddings, p_e_m=None)

Local entity scores

Returns:

  • –

    Entity scores.

Source code in /home/docs/checkouts/readthedocs.org/user_builds/rel/envs/latest/lib/python3.7/site-packages/REL/mulrel_ranker.py 
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
def __local_ent_scores(
    self, token_ids, tok_mask, entity_ids, entity_mask, embeddings, p_e_m=None
):
    """
    Local entity scores

    :return: Entity scores.
    """

    batchsize, n_words = token_ids.size()
    n_entities = entity_ids.size(1)
    tok_mask = tok_mask.view(batchsize, 1, -1)

    tok_vecs = embeddings["word_embeddings"](token_ids)
    entity_vecs = embeddings["entity_embeddings"](entity_ids)

    ent_tok_att_scores = torch.bmm(
        entity_vecs * self.att_mat_diag, tok_vecs.permute(0, 2, 1)
    )
    ent_tok_att_scores = (ent_tok_att_scores * tok_mask).add_(
        (tok_mask - 1).mul_(1e10)
    )
    tok_att_scores, _ = torch.max(ent_tok_att_scores, dim=1)
    top_tok_att_scores, top_tok_att_ids = torch.topk(
        tok_att_scores, dim=1, k=min(self.config["tok_top_n"], n_words)
    )
    att_probs = F.softmax(top_tok_att_scores, dim=1).view(batchsize, -1, 1)
    att_probs = att_probs / torch.sum(att_probs, dim=1, keepdim=True)

    selected_tok_vecs = torch.gather(
        tok_vecs,
        dim=1,
        index=top_tok_att_ids.view(batchsize, -1, 1).repeat(1, 1, tok_vecs.size(2)),
    )
    ctx_vecs = torch.sum(
        (selected_tok_vecs * self.tok_score_mat_diag) * att_probs,
        dim=1,
        keepdim=True,
    )
    ent_ctx_scores = torch.bmm(entity_vecs, ctx_vecs.permute(0, 2, 1)).view(
        batchsize, n_entities
    )

    # combine with p(e|m) if p_e_m is not None
    if p_e_m is not None:
        inputs = torch.cat(
            [
                ent_ctx_scores.view(batchsize * n_entities, -1),
                torch.log(p_e_m + 1e-20).view(batchsize * n_entities, -1),
            ],
            dim=1,
        )
        hidden = self.score_combine_linear_1(inputs)
        hidden = self.score_combine_act_1(hidden)
        scores = self.score_combine_linear_2(hidden).view(batchsize, n_entities)
    else:
        scores = ent_ctx_scores

    scores = (scores * entity_mask).add_((entity_mask - 1).mul_(1e10))

    self._entity_vecs = entity_vecs
    self._local_ctx_vecs = ctx_vecs

    return scores

forward(token_ids, tok_mask, entity_ids, entity_mask, p_e_m, embeddings, gold=None)

Responsible for forward pass of ED model and produces a ranking of candidates for a given set of mentions.

  • ctx_layer refers to function f. See Figure 3 in respective paper.
  • ent_scores refers to function q.
  • score_combine refers to function g.

Returns:

  • –

    Ranking of entities per mention.

Source code in /home/docs/checkouts/readthedocs.org/user_builds/rel/envs/latest/lib/python3.7/site-packages/REL/mulrel_ranker.py 
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
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
def forward(
    self, token_ids, tok_mask, entity_ids, entity_mask, p_e_m, embeddings, gold=None
):
    """
    Responsible for forward pass of ED model and produces a ranking of candidates for a given set of mentions.

    - ctx_layer refers to function f. See Figure 3 in respective paper.
    - ent_scores refers to function q.
    - score_combine refers to function g.

    :return: Ranking of entities per mention.
    """

    n_ments, n_cands = entity_ids.size()
    n_rels = self.config["n_rels"]

    if self.config["use_local"]:
        local_ent_scores = self.__local_ent_scores(
            token_ids, tok_mask, entity_ids, entity_mask, embeddings, p_e_m=None
        )
        ent_vecs = self._entity_vecs
    else:
        ent_vecs = embeddings["entity_embeddings"](entity_ids)
        local_ent_scores = Variable(
            torch.zeros(n_ments, n_cands), requires_grad=False
        ).to(self.device)

    # compute context vectors
    ltok_vecs = embeddings["snd_embeddings"](
        self.s_ltoken_ids
    ) * self.s_ltoken_mask.view(n_ments, -1, 1)
    local_lctx_vecs = torch.sum(ltok_vecs, dim=1) / torch.sum(
        self.s_ltoken_mask, dim=1, keepdim=True
    ).add_(1e-5)
    rtok_vecs = embeddings["snd_embeddings"](
        self.s_rtoken_ids
    ) * self.s_rtoken_mask.view(n_ments, -1, 1)
    local_rctx_vecs = torch.sum(rtok_vecs, dim=1) / torch.sum(
        self.s_rtoken_mask, dim=1, keepdim=True
    ).add_(1e-5)
    mtok_vecs = embeddings["snd_embeddings"](
        self.s_mtoken_ids
    ) * self.s_mtoken_mask.view(n_ments, -1, 1)
    ment_vecs = torch.sum(mtok_vecs, dim=1) / torch.sum(
        self.s_mtoken_mask, dim=1, keepdim=True
    ).add_(1e-5)
    bow_ctx_vecs = torch.cat([local_lctx_vecs, ment_vecs, local_rctx_vecs], dim=1)

    if self.config["use_pad_ent"]:
        ent_vecs = torch.cat(
            [ent_vecs, self.pad_ent_emb.view(1, 1, -1).repeat(1, n_cands, 1)], dim=0
        )
        tmp = torch.zeros(1, n_cands)
        tmp[0, 0] = 1
        tmp = Variable(tmp).to(self.device)
        entity_mask = torch.cat([entity_mask, tmp], dim=0)
        p_e_m = torch.cat([p_e_m, tmp], dim=0)
        local_ent_scores = torch.cat(
            [
                local_ent_scores,
                Variable(torch.zeros(1, n_cands), requires_grad=False).to(
                    self.device
                ),
            ],
            dim=0,
        )
        n_ments += 1

    if self.config["use_local_only"]:
        inputs = torch.cat(
            [
                Variable(torch.zeros(n_ments * n_cands, 1)).to(self.device),
                local_ent_scores.view(n_ments * n_cands, -1),
                torch.log(p_e_m + 1e-20).view(n_ments * n_cands, -1),
            ],
            dim=1,
        )
        scores = self.score_combine(inputs).view(n_ments, n_cands)
        return scores

    if n_ments == 1:
        ent_scores = local_ent_scores

    else:
        # distance - to consider only neighbor mentions
        ment_pos = torch.arange(0, n_ments).long()
        dist = (ment_pos.view(n_ments, 1) - ment_pos.view(1, n_ments)).abs()
        dist.masked_fill_(dist == 1, -1)
        dist.masked_fill_((dist > 1) & (dist <= self.max_dist), -1)
        dist.masked_fill_(dist > self.max_dist, 0)
        dist.mul_(-1)

        ctx_vecs = self.ctx_layer(bow_ctx_vecs)
        if self.config["use_pad_ent"]:
            ctx_vecs = torch.cat([ctx_vecs, self.pad_ctx_vec], dim=0)

        m1_ctx_vecs, m2_ctx_vecs = ctx_vecs, ctx_vecs
        rel_ctx_vecs = m1_ctx_vecs.view(1, n_ments, -1) * self.ew_embs.view(
            n_rels, 1, -1
        )
        rel_ctx_ctx_scores = torch.matmul(
            rel_ctx_vecs, m2_ctx_vecs.view(1, n_ments, -1).permute(0, 2, 1)
        )  # n_rels x n_ments x n_ments

        rel_ctx_ctx_scores = rel_ctx_ctx_scores.add_(
            (1 - Variable(dist.float()).to(self.device)).mul_(-1e10)
        )
        eye = Variable(torch.eye(n_ments)).view(1, n_ments, n_ments).to(self.device)
        rel_ctx_ctx_scores.add_(eye.mul_(-1e10))
        rel_ctx_ctx_scores.mul_(
            1 / np.sqrt(self.config["emb_dims"])
        )  # scaling proposed by "attention is all you need"

        # get top_n neighbour
        if self.ent_top_n < n_ments:
            topk_values, _ = torch.topk(
                rel_ctx_ctx_scores, k=min(self.ent_top_n, n_ments), dim=2
            )
            threshold = topk_values[:, :, -1:]
            mask = 1 - (rel_ctx_ctx_scores >= threshold).float()
            rel_ctx_ctx_scores.add_(mask.mul_(-1e10))

        rel_ctx_ctx_probs = F.softmax(rel_ctx_ctx_scores, dim=2)
        rel_ctx_ctx_weights = rel_ctx_ctx_probs + rel_ctx_ctx_probs.permute(0, 2, 1)
        self._rel_ctx_ctx_weights = rel_ctx_ctx_probs

        # compute phi(ei, ej)
        rel_ent_vecs = ent_vecs.view(1, n_ments, n_cands, -1) * self.rel_embs.view(
            n_rels, 1, 1, -1
        )
        rel_ent_ent_scores = torch.matmul(
            rel_ent_vecs.view(n_rels, n_ments, 1, n_cands, -1),
            ent_vecs.view(1, 1, n_ments, n_cands, -1).permute(0, 1, 2, 4, 3),
        )

        rel_ent_ent_scores = rel_ent_ent_scores.permute(
            0, 1, 3, 2, 4
        )  # n_rel x n_ments x n_cands x n_ments x n_cands
        rel_ent_ent_scores = (rel_ent_ent_scores * entity_mask).add_(
            (entity_mask - 1).mul_(1e10)
        )
        ent_ent_scores = torch.sum(
            rel_ent_ent_scores
            * rel_ctx_ctx_weights.view(n_rels, n_ments, 1, n_ments, 1),
            dim=0,
        ).mul(
            1.0 / n_rels
        )  # n_ments x n_cands x n_ments x n_cands

        # LBP
        prev_msgs = Variable(torch.zeros(n_ments, n_cands, n_ments)).to(self.device)

        for _ in range(self.config["n_loops"]):
            mask = 1 - Variable(torch.eye(n_ments)).to(self.device)
            ent_ent_votes = (
                ent_ent_scores
                + local_ent_scores * 1
                + torch.sum(
                    prev_msgs.view(1, n_ments, n_cands, n_ments)
                    * mask.view(n_ments, 1, 1, n_ments),
                    dim=3,
                ).view(n_ments, 1, n_ments, n_cands)
            )
            msgs, _ = torch.max(ent_ent_votes, dim=3)
            msgs = (
                F.softmax(msgs, dim=1).mul(self.config["dropout_rate"])
                + prev_msgs.exp().mul(1 - self.config["dropout_rate"])
            ).log()
            prev_msgs = msgs

        # compute marginal belief
        mask = 1 - Variable(torch.eye(n_ments)).to(self.device)
        ent_scores = local_ent_scores * 1 + torch.sum(
            msgs * mask.view(n_ments, 1, n_ments), dim=2
        )
        ent_scores = F.softmax(ent_scores, dim=1)

    # combine with p_e_m
    inputs = torch.cat(
        [
            ent_scores.view(n_ments * n_cands, -1),
            torch.log(p_e_m + 1e-20).view(n_ments * n_cands, -1),
        ],
        dim=1,
    )
    scores = self.score_combine(inputs).view(n_ments, n_cands)
    # scores = F.softmax(scores, dim=1)

    if self.config["use_pad_ent"]:
        scores = scores[:-1]
    return scores, ent_scores

loss(scores, true_pos, lamb=1e-07)

Computes given ranking loss (Equation 7) and adds a regularization term.

Returns:

  • –

    loss of given batch

Source code in /home/docs/checkouts/readthedocs.org/user_builds/rel/envs/latest/lib/python3.7/site-packages/REL/mulrel_ranker.py 
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
def loss(self, scores, true_pos, lamb=1e-7):
    """
    Computes given ranking loss (Equation 7) and adds a regularization term.

    :return: loss of given batch
    """
    loss = F.multi_margin_loss(scores, true_pos, margin=self.config["margin"])
    if self.config["use_local_only"]:
        return loss

    # regularization
    X = F.normalize(self.rel_embs)
    diff = (
        (
            X.view(self.config["n_rels"], 1, -1)
            - X.view(1, self.config["n_rels"], -1)
        )
        .pow(2)
        .sum(dim=2)
        .add_(1e-5)
        .sqrt()
    )
    diff = diff * (diff < 1).float()
    loss -= torch.sum(diff).mul(lamb)

    X = F.normalize(self.ew_embs)
    diff = (
        (
            X.view(self.config["n_rels"], 1, -1)
            - X.view(1, self.config["n_rels"], -1)
        )
        .pow(2)
        .sum(dim=2)
        .add_(1e-5)
        .sqrt()
    )
    diff = diff * (diff < 1).float()
    loss -= torch.sum(diff).mul(lamb)
    return loss

regularize(max_norm=1)

Regularises model parameters.

Returns:

  • –

    -

Source code in /home/docs/checkouts/readthedocs.org/user_builds/rel/envs/latest/lib/python3.7/site-packages/REL/mulrel_ranker.py 
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
def regularize(self, max_norm=1):
    """
    Regularises model parameters.

    :return: -
    """

    l1_w_norm = self.score_combine_linear_1.weight.norm()
    l1_b_norm = self.score_combine_linear_1.bias.norm()
    l2_w_norm = self.score_combine_linear_2.weight.norm()
    l2_b_norm = self.score_combine_linear_2.bias.norm()

    if (l1_w_norm > max_norm).data.all():
        self.score_combine_linear_1.weight.data = (
            self.score_combine_linear_1.weight.data * max_norm / l1_w_norm.data
        )
    if (l1_b_norm > max_norm).data.all():
        self.score_combine_linear_1.bias.data = (
            self.score_combine_linear_1.bias.data * max_norm / l1_b_norm.data
        )
    if (l2_w_norm > max_norm).data.all():
        self.score_combine_linear_2.weight.data = (
            self.score_combine_linear_2.weight.data * max_norm / l2_w_norm.data
        )
    if (l2_b_norm > max_norm).data.all():
        self.score_combine_linear_2.bias.data = (
            self.score_combine_linear_2.bias.data * max_norm / l2_b_norm.data
        )

PreRank(config, embeddings=None)

Bases: torch.nn.Module

PreRank class is used for preranking entities for a given mention by multiplying entity vectors with word vectors

Source code in /home/docs/checkouts/readthedocs.org/user_builds/rel/envs/latest/lib/python3.7/site-packages/REL/mulrel_ranker.py 
13
14
15
def __init__(self, config, embeddings=None):
    super(PreRank, self).__init__()
    self.config = config

forward(token_ids, token_offsets, entity_ids, embeddings, emb)

Multiplies local context words with entity vectors for a given mention.

Returns:

  • –

    entity scores.

Source code in /home/docs/checkouts/readthedocs.org/user_builds/rel/envs/latest/lib/python3.7/site-packages/REL/mulrel_ranker.py 
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
def forward(self, token_ids, token_offsets, entity_ids, embeddings, emb):
    """
    Multiplies local context words with entity vectors for a given mention.

    :return: entity scores.
    """

    sent_vecs = embeddings["word_embeddings_bag"](
        token_ids, token_offsets
    )  # (batch_size, emb_size=300)

    # entity_vecs = emb.emb(entity_names)

    entity_vecs = embeddings["entity_embeddings"](
        entity_ids
    )  # (batch_size, n_cands, emb_size)

    # compute scores
    batchsize, dims = sent_vecs.size()
    n_entities = entity_vecs.size(1)
    scores = torch.bmm(entity_vecs, sent_vecs.view(batchsize, dims, 1))
    scores = scores.view(batchsize, n_entities)

    log_probs = F.log_softmax(scores, dim=1)
    return log_probs