def forward(
self, # type: ignore
text: Dict[str, torch.LongTensor],
spans: torch.IntTensor,
span_labels: torch.IntTensor = None,
metadata: List[Dict[str, Any]] = None,
) -> Dict[str, torch.Tensor]:
# pylint: disable=arguments-differ
# Shape: (batch_size, document_length, embedding_size)
text_embeddings = self._lexical_dropout(
self._text_field_embedder(text))
document_length = text_embeddings.size(1)
# Shape: (batch_size, document_length)
text_mask = util.get_text_field_mask(text).float()
# Shape: (batch_size, num_spans)
if self._use_gold_mentions:
if text_embeddings.is_cuda:
device = torch.device("cuda")
else:
device = torch.device("cpu")
s = [
torch.as_tensor(pair, dtype=torch.long, device=device)
for cluster in metadata[0]["clusters"] for pair in cluster
]
gm = torch.stack(s, dim=0).unsqueeze(0).unsqueeze(1)
span_mask = spans.unsqueeze(2) - gm
span_mask = (span_mask[:, :, :, 0] == 0) + (span_mask[:, :, :, 1]
== 0)
span_mask, _ = (span_mask == 2).max(-1)
num_spans = span_mask.sum().item()
span_mask = span_mask.float()
else:
span_mask = (spans[:, :, 0] >= 0).squeeze(-1).float()
num_spans = spans.size(1)
# Shape: (batch_size, num_spans, 2)
spans = F.relu(spans.float()).long()
# Shape: (batch_size, document_length, encoding_dim)
contextualized_embeddings = self._context_layer(
text_embeddings, text_mask)
# Shape: (batch_size, num_spans, 2 * encoding_dim + feature_size)
endpoint_span_embeddings = self._endpoint_span_extractor(
contextualized_embeddings, spans)
# Shape: (batch_size, num_spans, emebedding_size)
attended_span_embeddings = self._attentive_span_extractor(
text_embeddings, spans)
# Shape: (batch_size, num_spans, emebedding_size + 2 * encoding_dim + feature_size)
span_embeddings = torch.cat(
[endpoint_span_embeddings, attended_span_embeddings], -1)
# Prune based on mention scores.
num_spans_to_keep = int(
math.floor(self._spans_per_word * document_length))
(top_span_embeddings, top_span_mask, top_span_indices,
top_span_mention_scores) = self._mention_pruner(
span_embeddings, span_mask, num_spans_to_keep)
top_span_mask = top_span_mask.unsqueeze(-1)
# Shape: (batch_size * num_spans_to_keep)
flat_top_span_indices = util.flatten_and_batch_shift_indices(
top_span_indices, num_spans)
# Compute final predictions for which spans to consider as mentions.
# Shape: (batch_size, num_spans_to_keep, 2)
top_spans = util.batched_index_select(spans, top_span_indices,
flat_top_span_indices)
# Compute indices for antecedent spans to consider.
max_antecedents = min(self._max_antecedents, num_spans_to_keep)
# Shapes:
# (num_spans_to_keep, max_antecedents),
# (1, max_antecedents),
# (1, num_spans_to_keep, max_antecedents)
valid_antecedent_indices, valid_antecedent_offsets, valid_antecedent_log_mask = self._generate_valid_antecedents(
num_spans_to_keep, max_antecedents, util.get_device_of(text_mask))
# Select tensors relating to the antecedent spans.
# Shape: (batch_size, num_spans_to_keep, max_antecedents, embedding_size)
candidate_antecedent_embeddings = util.flattened_index_select(
top_span_embeddings, valid_antecedent_indices)
# Shape: (batch_size, num_spans_to_keep, max_antecedents)
candidate_antecedent_mention_scores = util.flattened_index_select(
top_span_mention_scores, valid_antecedent_indices).squeeze(-1)
# Compute antecedent scores.
# Shape: (batch_size, num_spans_to_keep, max_antecedents, embedding_size)
span_pair_embeddings = self._compute_span_pair_embeddings(
top_span_embeddings, candidate_antecedent_embeddings,
valid_antecedent_offsets)
# Shape: (batch_size, num_spans_to_keep, 1 + max_antecedents)
coreference_scores = self._compute_coreference_scores(
span_pair_embeddings,
top_span_mention_scores,
candidate_antecedent_mention_scores,
valid_antecedent_log_mask,
)
# Shape: (batch_size, num_spans_to_keep)
_, predicted_antecedents = coreference_scores.max(2)
predicted_antecedents -= 1
output_dict = {
"top_spans": top_spans,
"antecedent_indices": valid_antecedent_indices,
"predicted_antecedents": predicted_antecedents,
}
if span_labels is not None:
# Find the gold labels for the spans which we kept.
pruned_gold_labels = util.batched_index_select(
span_labels.unsqueeze(-1), top_span_indices,
flat_top_span_indices)
antecedent_labels = util.flattened_index_select(
pruned_gold_labels, valid_antecedent_indices).squeeze(-1)
antecedent_labels += valid_antecedent_log_mask.long()
# Compute labels.
# Shape: (batch_size, num_spans_to_keep, max_antecedents + 1)
gold_antecedent_labels = self._compute_antecedent_gold_labels(
pruned_gold_labels, antecedent_labels)
coreference_log_probs = util.last_dim_log_softmax(
coreference_scores, top_span_mask)
correct_antecedent_log_probs = coreference_log_probs + gold_antecedent_labels.log(
)
negative_marginal_log_likelihood = -util.logsumexp(
correct_antecedent_log_probs).sum()
self._mention_recall(top_spans, metadata)
self._conll_coref_scores(top_spans, valid_antecedent_indices,
predicted_antecedents, metadata)
output_dict["loss"] = negative_marginal_log_likelihood
if metadata is not None:
output_dict["document"] = [x["original_text"] for x in metadata]
return output_dict
def _construct_loss(
self, head_tag_representation: torch.Tensor,
child_tag_representation: torch.Tensor,
attended_arcs: torch.Tensor, head_indices: torch.Tensor,
head_tags: torch.Tensor,
mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Computes the arc and tag loss for a sequence given gold head indices and tags.
Parameters
----------
head_tag_representation : ``torch.Tensor``, required.
A tensor of shape (batch_size, sequence_length, tag_representation_dim),
which will be used to generate predictions for the dependency tags
for the given arcs.
child_tag_representation : ``torch.Tensor``, required
A tensor of shape (batch_size, sequence_length, tag_representation_dim),
which will be used to generate predictions for the dependency tags
for the given arcs.
attended_arcs : ``torch.Tensor``, required.
A tensor of shape (batch_size, sequence_length, sequence_length) used to generate
a distribution over attachements of a given word to all other words.
head_indices : ``torch.Tensor``, required.
A tensor of shape (batch_size, sequence_length).
The indices of the heads for every word.
head_tags : ``torch.Tensor``, required.
A tensor of shape (batch_size, sequence_length).
The dependency labels of the heads for every word.
mask : ``torch.Tensor``, required.
A mask of shape (batch_size, sequence_length), denoting unpadded
elements in the sequence.
Returns
-------
arc_nll : ``torch.Tensor``, required.
The negative log likelihood from the arc loss.
tag_nll : ``torch.Tensor``, required.
The negative log likelihood from the arc tag loss.
"""
float_mask = mask.float()
batch_size, sequence_length, _ = attended_arcs.size()
# shape (batch_size, 1)
range_vector = get_range_vector(
batch_size, get_device_of(attended_arcs)).unsqueeze(1)
# shape (batch_size, sequence_length, sequence_length)
normalised_arc_logits = last_dim_log_softmax(
attended_arcs,
mask) * float_mask.unsqueeze(2) * float_mask.unsqueeze(1)
# shape (batch_size, sequence_length, num_head_tags)
head_tag_logits = self._get_head_tags(head_tag_representation,
child_tag_representation,
head_indices)
normalised_head_tag_logits = last_dim_log_softmax(
head_tag_logits, mask.unsqueeze(-1)) * float_mask.unsqueeze(-1)
# index matrix with shape (batch, sequence_length)
timestep_index = get_range_vector(sequence_length,
get_device_of(attended_arcs))
child_index = timestep_index.view(1, sequence_length).expand(
batch_size, sequence_length).long()
# shape (batch_size, sequence_length)
arc_loss = normalised_arc_logits[range_vector, child_index,
head_indices]
tag_loss = normalised_head_tag_logits[range_vector, child_index,
head_tags]
# We don't care about predictions for the symbolic ROOT token's head,
# so we remove it from the loss.
arc_loss = arc_loss[:, 1:]
tag_loss = tag_loss[:, 1:]
# The number of valid positions is equal to the number of unmasked elements minus
# 1 per sequence in the batch, to account for the symbolic HEAD token.
valid_positions = mask.sum() - batch_size
arc_nll = -arc_loss.sum() / valid_positions.float()
tag_nll = -tag_loss.sum() / valid_positions.float()
return arc_nll, tag_nll
def forward(self, # type: ignore
text: Dict[str, torch.LongTensor],
spans: torch.IntTensor,
span_labels: torch.IntTensor = None,
metadata: List[Dict[str, Any]] = None) -> Dict[str, torch.Tensor]:
# pylint: disable=arguments-differ
"""
Parameters
----------
text : ``Dict[str, torch.LongTensor]``, required.
The output of a ``TextField`` representing the text of
the document.
spans : ``torch.IntTensor``, required.
A tensor of shape (batch_size, num_spans, 2), representing the inclusive start and end
indices of candidate spans for mentions. Comes from a ``ListField[SpanField]`` of
indices into the text of the document.
span_labels : ``torch.IntTensor``, optional (default = None)
A tensor of shape (batch_size, num_spans), representing the cluster ids
of each span, or -1 for those which do not appear in any clusters.
Returns
-------
An output dictionary consisting of:
top_spans : ``torch.IntTensor``
A tensor of shape ``(batch_size, num_spans_to_keep, 2)`` representing
the start and end word indices of the top spans that survived the pruning stage.
antecedent_indices : ``torch.IntTensor``
A tensor of shape ``(num_spans_to_keep, max_antecedents)`` representing for each top span
the index (with respect to top_spans) of the possible antecedents the model considered.
predicted_antecedents : ``torch.IntTensor``
A tensor of shape ``(batch_size, num_spans_to_keep)`` representing, for each top span, the
index (with respect to antecedent_indices) of the most likely antecedent. -1 means there
was no predicted link.
loss : ``torch.FloatTensor``, optional
A scalar loss to be optimised.
"""
# Shape: (batch_size, document_length, embedding_size)
text_embeddings = self._lexical_dropout(self._text_field_embedder(text))
document_length = text_embeddings.size(1)
num_spans = spans.size(1)
# Shape: (batch_size, document_length)
text_mask = util.get_text_field_mask(text).float()
# Shape: (batch_size, num_spans)
span_mask = (spans[:, :, 0] >= 0).squeeze(-1).float()
# SpanFields return -1 when they are used as padding. As we do
# some comparisons based on span widths when we attend over the
# span representations that we generate from these indices, we
# need them to be <= 0. This is only relevant in edge cases where
# the number of spans we consider after the pruning stage is >= the
# total number of spans, because in this case, it is possible we might
# consider a masked span.
# Shape: (batch_size, num_spans, 2)
spans = F.relu(spans.float()).long()
# Shape: (batch_size, document_length, encoding_dim)
contextualized_embeddings = self._context_layer(text_embeddings, text_mask)
# Shape: (batch_size, num_spans, 2 * encoding_dim + feature_size)
endpoint_span_embeddings = self._endpoint_span_extractor(contextualized_embeddings, spans)
# Shape: (batch_size, num_spans, emebedding_size)
attended_span_embeddings = self._attentive_span_extractor(text_embeddings, spans)
# Shape: (batch_size, num_spans, emebedding_size + 2 * encoding_dim + feature_size)
span_embeddings = torch.cat([endpoint_span_embeddings, attended_span_embeddings], -1)
# Prune based on mention scores.
num_spans_to_keep = int(math.floor(self._spans_per_word * document_length))
(top_span_embeddings, top_span_mask,
top_span_indices, top_span_mention_scores) = self._mention_pruner(span_embeddings,
span_mask,
num_spans_to_keep)
top_span_mask = top_span_mask.unsqueeze(-1)
# Shape: (batch_size * num_spans_to_keep)
# torch.index_select only accepts 1D indices, but here
# we need to select spans for each element in the batch.
# This reformats the indices to take into account their
# index into the batch. We precompute this here to make
# the multiple calls to util.batched_index_select below more efficient.
flat_top_span_indices = util.flatten_and_batch_shift_indices(top_span_indices, num_spans)
# Compute final predictions for which spans to consider as mentions.
# Shape: (batch_size, num_spans_to_keep, 2)
top_spans = util.batched_index_select(spans,
top_span_indices,
flat_top_span_indices)
# Compute indices for antecedent spans to consider.
max_antecedents = min(self._max_antecedents, num_spans_to_keep)
# Now that we have our variables in terms of num_spans_to_keep, we need to
# compare span pairs to decide each span's antecedent. Each span can only
# have prior spans as antecedents, and we only consider up to max_antecedents
# prior spans. So the first thing we do is construct a matrix mapping a span's
# index to the indices of its allowed antecedents. Note that this is independent
# of the batch dimension - it's just a function of the span's position in
# top_spans. The spans are in document order, so we can just use the relative
# index of the spans to know which other spans are allowed antecedents.
# Once we have this matrix, we reformat our variables again to get embeddings
# for all valid antecedents for each span. This gives us variables with shapes
# like (batch_size, num_spans_to_keep, max_antecedents, embedding_size), which
# we can use to make coreference decisions between valid span pairs.
# Shapes:
# (num_spans_to_keep, max_antecedents),
# (1, max_antecedents),
# (1, num_spans_to_keep, max_antecedents)
valid_antecedent_indices, valid_antecedent_offsets, valid_antecedent_log_mask = \
self._generate_valid_antecedents(num_spans_to_keep, max_antecedents, util.get_device_of(text_mask))
# Select tensors relating to the antecedent spans.
# Shape: (batch_size, num_spans_to_keep, max_antecedents, embedding_size)
candidate_antecedent_embeddings = util.flattened_index_select(top_span_embeddings,
valid_antecedent_indices)
# Shape: (batch_size, num_spans_to_keep, max_antecedents)
candidate_antecedent_mention_scores = util.flattened_index_select(top_span_mention_scores,
valid_antecedent_indices).squeeze(-1)
# Compute antecedent scores.
# Shape: (batch_size, num_spans_to_keep, max_antecedents, embedding_size)
span_pair_embeddings = self._compute_span_pair_embeddings(top_span_embeddings,
candidate_antecedent_embeddings,
valid_antecedent_offsets)
# Shape: (batch_size, num_spans_to_keep, 1 + max_antecedents)
coreference_scores = self._compute_coreference_scores(span_pair_embeddings,
top_span_mention_scores,
candidate_antecedent_mention_scores,
valid_antecedent_log_mask)
# We now have, for each span which survived the pruning stage,
# a predicted antecedent. This implies a clustering if we group
# mentions which refer to each other in a chain.
# Shape: (batch_size, num_spans_to_keep)
_, predicted_antecedents = coreference_scores.max(2)
# Subtract one here because index 0 is the "no antecedent" class,
# so this makes the indices line up with actual spans if the prediction
# is greater than -1.
predicted_antecedents -= 1
output_dict = {"top_spans": top_spans,
"antecedent_indices": valid_antecedent_indices,
"predicted_antecedents": predicted_antecedents}
if span_labels is not None:
# Find the gold labels for the spans which we kept.
pruned_gold_labels = util.batched_index_select(span_labels.unsqueeze(-1),
top_span_indices,
flat_top_span_indices)
antecedent_labels = util.flattened_index_select(pruned_gold_labels,
valid_antecedent_indices).squeeze(-1)
antecedent_labels += valid_antecedent_log_mask.long()
# Compute labels.
# Shape: (batch_size, num_spans_to_keep, max_antecedents + 1)
gold_antecedent_labels = self._compute_antecedent_gold_labels(pruned_gold_labels,
antecedent_labels)
# Now, compute the loss using the negative marginal log-likelihood.
# This is equal to the log of the sum of the probabilities of all antecedent predictions
# that would be consistent with the data, in the sense that we are minimising, for a
# given span, the negative marginal log likelihood of all antecedents which are in the
# same gold cluster as the span we are currently considering. Each span i predicts a
# single antecedent j, but there might be several prior mentions k in the same
# coreference cluster that would be valid antecedents. Our loss is the sum of the
# probability assigned to all valid antecedents. This is a valid objective for
# clustering as we don't mind which antecedent is predicted, so long as they are in
# the same coreference cluster.
coreference_log_probs = util.last_dim_log_softmax(coreference_scores, top_span_mask)
correct_antecedent_log_probs = coreference_log_probs + gold_antecedent_labels.log()
negative_marginal_log_likelihood = -util.logsumexp(correct_antecedent_log_probs).sum()
self._mention_recall(top_spans, metadata)
self._conll_coref_scores(top_spans, valid_antecedent_indices, predicted_antecedents, metadata)
output_dict["loss"] = negative_marginal_log_likelihood
return output_dict
def _construct_loss(self,
head_tag_representation: torch.Tensor,
child_tag_representation: torch.Tensor,
attended_arcs: torch.Tensor,
head_indices: torch.Tensor,
head_tags: torch.Tensor,
mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Computes the arc and tag loss for a sequence given gold head indices and tags.
Parameters
----------
head_tag_representation : ``torch.Tensor``, required.
A tensor of shape (batch_size, sequence_length, tag_representation_dim),
which will be used to generate predictions for the dependency tags
for the given arcs.
child_tag_representation : ``torch.Tensor``, required
A tensor of shape (batch_size, sequence_length, tag_representation_dim),
which will be used to generate predictions for the dependency tags
for the given arcs.
attended_arcs : ``torch.Tensor``, required.
A tensor of shape (batch_size, sequence_length, sequence_length) used to generate
a distribution over attachements of a given word to all other words.
head_indices : ``torch.Tensor``, required.
A tensor of shape (batch_size, sequence_length).
The indices of the heads for every word.
head_tags : ``torch.Tensor``, required.
A tensor of shape (batch_size, sequence_length).
The dependency labels of the heads for every word.
mask : ``torch.Tensor``, required.
A mask of shape (batch_size, sequence_length), denoting unpadded
elements in the sequence.
Returns
-------
arc_nll : ``torch.Tensor``, required.
The negative log likelihood from the arc loss.
tag_nll : ``torch.Tensor``, required.
The negative log likelihood from the arc tag loss.
"""
float_mask = mask.float()
batch_size, sequence_length, _ = attended_arcs.size()
# shape (batch_size, 1)
range_vector = get_range_vector(batch_size, get_device_of(attended_arcs)).unsqueeze(1)
# shape (batch_size, sequence_length, sequence_length)
normalised_arc_logits = last_dim_log_softmax(attended_arcs,
mask) * float_mask.unsqueeze(2) * float_mask.unsqueeze(1)
# shape (batch_size, sequence_length, num_head_tags)
head_tag_logits = self._get_head_tags(head_tag_representation, child_tag_representation, head_indices)
normalised_head_tag_logits = last_dim_log_softmax(head_tag_logits,
mask.unsqueeze(-1)) * float_mask.unsqueeze(-1)
# index matrix with shape (batch, sequence_length)
timestep_index = get_range_vector(sequence_length, get_device_of(attended_arcs))
child_index = timestep_index.view(1, sequence_length).expand(batch_size, sequence_length).long()
# shape (batch_size, sequence_length)
arc_loss = normalised_arc_logits[range_vector, child_index, head_indices]
tag_loss = normalised_head_tag_logits[range_vector, child_index, head_tags]
# We don't care about predictions for the symbolic ROOT token's head,
# so we remove it from the loss.
arc_loss = arc_loss[:, 1:]
tag_loss = tag_loss[:, 1:]
# The number of valid positions is equal to the number of unmasked elements minus
# 1 per sequence in the batch, to account for the symbolic HEAD token.
valid_positions = mask.sum() - batch_size
arc_nll = -arc_loss.sum() / valid_positions.float()
tag_nll = -tag_loss.sum() / valid_positions.float()
return arc_nll, tag_nll
