target_extraction.allen.modules.target_position_weight package¶

Submodules¶

target_extraction.allen.modules.target_position_weight.relative_target_position_weight module¶

class target_extraction.allen.modules.target_position_weight.relative_target_position_weight.RelativeTargetPositionWeight(zero_target_word_weighting=False)[source]¶

Bases: target_extraction.allen.modules.target_position_weight.target_position_weight.TargetPositionWeight

The weighting performed here is the following:

$$1 -

rac{| heta - i|}{n}$$

Where $i$ is the location of the token, $ heta$ is the location of the nearest target token (can be more than one taregt token in the sentence if the target is a multi-word target), and $n$ is the token length of the text. The weight of the target tokens by default is $1$ thus target tokens are not down weighted. This is the same weighting as equation 7 within Chen et al. 2017 and equation 2 in Zhao et al. 2019

param zero_target_word_weighting

If True it will apply a weight of 0 to all target words (same as masking the target words). This would be the same weighting function as Zhang et al. 2019

forward(targets_features, relative_target_positions, sequence_mask)[source]¶

Parameters

targets_features (Tensor) – A tensor of shape (batch * num_targets, text_sequence_length, dim). This tensor will be returned weighted by the position of the tokens in the sequence with respect to the target tokens.
relative_target_positions (Tensor) – A tensor of shape (batch, num_targets, text_sequence_length). This will be a tensor that contains the position of each token to its associated target tokens in the sample.
sequence_mask (Tensor) – A tensor of shape (batch * num_targets, text_sequence_length). The mask determines which tokens are to be weighted based on their position in the sequence.

Return type

Tuple[Tensor, Tensor]

Returns

A tuple of two tensors 1. tensor of shape (batch * num_targets, text_sequence_length, dim), where the target_features have been weighted based on each tokens position to its sample’s respective target token position. 2. tensor of shape (batch * num_targets, text_sequence_length) representing the weights that the target_features have been multipled by to get the first tensor in this tuple.

Raises

ConfigurationError – If the targets_features first dimension is not batch size * num targets size.

target_extraction.allen.modules.target_position_weight.target_position_weight module¶

class target_extraction.allen.modules.target_position_weight.target_position_weight.TargetPositionWeight[source]¶

Bases: torch.nn.modules.module.Module, allennlp.common.registrable.Registrable

A TargetPositionWeight is a Module that represents different methods that can weight a target sample’s encoded text by the position the tokens take in the text with respect to the target tokens.