# LUKE¶

## Overview¶

The LUKE model was proposed in LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda and Yuji Matsumoto. It is based on RoBERTa and adds entity embeddings as well as an entity-aware self-attention mechanism, which helps improve performance on various downstream tasks involving reasoning about entities such as named entity recognition, extractive and cloze-style question answering, entity typing, and relation classification.

The abstract from the paper is the following:

Tips:

• This implementation is the same as RobertaModel with the addition of entity embeddings as well as an entity-aware self-attention mechanism, which improves performance on tasks involving reasoning about entities.

• LUKE treats entities as input tokens; therefore, it takes entity_ids, entity_attention_mask, entity_token_type_ids and entity_position_ids as extra input. You can obtain those using LukeTokenizer.

• LukeTokenizer takes entities and entity_spans (character-based start and end positions of the entities in the input text) as extra input. entities typically consist of [MASK] entities or Wikipedia entities. The brief description when inputting these entities are as follows:

• Inputting [MASK] entities to compute entity representations: The [MASK] entity is used to mask entities to be predicted during pretraining. When LUKE receives the [MASK] entity, it tries to predict the original entity by gathering the information about the entity from the input text. Therefore, the [MASK] entity can be used to address downstream tasks requiring the information of entities in text such as entity typing, relation classification, and named entity recognition.

• Inputting Wikipedia entities to compute knowledge-enhanced token representations: LUKE learns rich information (or knowledge) about Wikipedia entities during pretraining and stores the information in its entity embedding. By using Wikipedia entities as input tokens, LUKE outputs token representations enriched by the information stored in the embeddings of these entities. This is particularly effective for tasks requiring real-world knowledge, such as question answering.

• There are three head models for the former use case:

• LukeForEntityClassification, for tasks to classify a single entity in an input text such as entity typing, e.g. the Open Entity dataset. This model places a linear head on top of the output entity representation.

• LukeForEntityPairClassification, for tasks to classify the relationship between two entities such as relation classification, e.g. the TACRED dataset. This model places a linear head on top of the concatenated output representation of the pair of given entities.

• LukeForEntitySpanClassification, for tasks to classify the sequence of entity spans, such as named entity recognition (NER). This model places a linear head on top of the output entity representations. You can address NER using this model by inputting all possible entity spans in the text to the model.

LukeTokenizer has a task argument, which enables you to easily create an input to these head models by specifying task="entity_classification", task="entity_pair_classification", or task="entity_span_classification". Please refer to the example code of each head models.

There are also 3 notebooks available, which showcase how you can reproduce the results as reported in the paper with the HuggingFace implementation of LUKE. They can be found here.

Example:

>>> from transformers import LukeTokenizer, LukeModel, LukeForEntityPairClassification

>>> model = LukeModel.from_pretrained("studio-ousia/luke-base")
>>> tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base")

# Example 1: Computing the contextualized entity representation corresponding to the entity mention "Beyoncé"
>>> text = "Beyoncé lives in Los Angeles."
>>> entity_spans = [(0, 7)]  # character-based entity span corresponding to "Beyoncé"
>>> inputs = tokenizer(text, entity_spans=entity_spans, add_prefix_space=True, return_tensors="pt")
>>> outputs = model(**inputs)
>>> word_last_hidden_state = outputs.last_hidden_state
>>> entity_last_hidden_state = outputs.entity_last_hidden_state

# Example 2: Inputting Wikipedia entities to obtain enriched contextualized representations
>>> entities = ["Beyoncé", "Los Angeles"]  # Wikipedia entity titles corresponding to the entity mentions "Beyoncé" and "Los Angeles"
>>> entity_spans = [(0, 7), (17, 28)]  # character-based entity spans corresponding to "Beyoncé" and "Los Angeles"
>>> inputs = tokenizer(text, entities=entities, entity_spans=entity_spans, add_prefix_space=True, return_tensors="pt")
>>> outputs = model(**inputs)
>>> word_last_hidden_state = outputs.last_hidden_state
>>> entity_last_hidden_state = outputs.entity_last_hidden_state

# Example 3: Classifying the relationship between two entities using LukeForEntityPairClassification head model
>>> model = LukeForEntityPairClassification.from_pretrained("studio-ousia/luke-large-finetuned-tacred")
>>> tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-large-finetuned-tacred")
>>> entity_spans = [(0, 7), (17, 28)]  # character-based entity spans corresponding to "Beyoncé" and "Los Angeles"
>>> inputs = tokenizer(text, entity_spans=entity_spans, return_tensors="pt")
>>> outputs = model(**inputs)
>>> logits = outputs.logits
>>> predicted_class_idx = int(logits[0].argmax())
>>> print("Predicted class:", model.config.id2label[predicted_class_idx])


This model was contributed by ikuyamada and nielsr. The original code can be found here.

## LukeConfig¶

class transformers.LukeConfig(vocab_size=50267, entity_vocab_size=500000, hidden_size=768, entity_emb_size=256, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, hidden_act='gelu', hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, type_vocab_size=2, initializer_range=0.02, layer_norm_eps=1e-12, gradient_checkpointing=False, use_entity_aware_attention=True, pad_token_id=1, bos_token_id=0, eos_token_id=2, **kwargs)[source]

This is the configuration class to store the configuration of a LukeModel. It is used to instantiate a LUKE model according to the specified arguments, defining the model architecture.

Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.

Parameters
• vocab_size (int, optional, defaults to 30522) – Vocabulary size of the LUKE model. Defines the number of different tokens that can be represented by the inputs_ids passed when calling LukeModel.

• entity_vocab_size (int, optional, defaults to 500000) – Entity vocabulary size of the LUKE model. Defines the number of different entities that can be represented by the entity_ids passed when calling LukeModel.

• hidden_size (int, optional, defaults to 768) – Dimensionality of the encoder layers and the pooler layer.

• entity_emb_size (int, optional, defaults to 256) – The number of dimensions of the entity embedding.

• num_hidden_layers (int, optional, defaults to 12) – Number of hidden layers in the Transformer encoder.

• num_attention_heads (int, optional, defaults to 12) – Number of attention heads for each attention layer in the Transformer encoder.

• intermediate_size (int, optional, defaults to 3072) – Dimensionality of the “intermediate” (often named feed-forward) layer in the Transformer encoder.

• hidden_act (str or Callable, optional, defaults to "gelu") – The non-linear activation function (function or string) in the encoder and pooler. If string, "gelu", "relu", "silu" and "gelu_new" are supported.

• hidden_dropout_prob (float, optional, defaults to 0.1) – The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.

• attention_probs_dropout_prob (float, optional, defaults to 0.1) – The dropout ratio for the attention probabilities.

• max_position_embeddings (int, optional, defaults to 512) – The maximum sequence length that this model might ever be used with. Typically set this to something large just in case (e.g., 512 or 1024 or 2048).

• type_vocab_size (int, optional, defaults to 2) – The vocabulary size of the token_type_ids passed when calling LukeModel.

• initializer_range (float, optional, defaults to 0.02) – The standard deviation of the truncated_normal_initializer for initializing all weight matrices.

• layer_norm_eps (float, optional, defaults to 1e-12) – The epsilon used by the layer normalization layers.

• gradient_checkpointing (bool, optional, defaults to False) – If True, use gradient checkpointing to save memory at the expense of slower backward pass.

• use_entity_aware_attention (bool, defaults to True) – Whether or not the model should use the entity-aware self-attention mechanism proposed in LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention (Yamada et al.).

Examples:

>>> from transformers import LukeConfig, LukeModel

>>> # Initializing a LUKE configuration
>>> configuration = LukeConfig()

>>> # Initializing a model from the configuration
>>> model = LukeModel(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config


## LukeTokenizer¶

class transformers.LukeTokenizer(vocab_file, merges_file, entity_vocab_file, task=None, max_entity_length=32, max_mention_length=30, entity_token_1='<ent>', entity_token_2='<ent2>', **kwargs)[source]

Construct a LUKE tokenizer.

This tokenizer inherits from RobertaTokenizer which contains most of the main methods. Users should refer to this superclass for more information regarding those methods. Compared to RobertaTokenizer, LukeTokenizer also creates entity sequences, namely entity_ids, entity_attention_mask, entity_token_type_ids, and entity_position_ids to be used by the LUKE model.

Parameters
• vocab_file (str) – Path to the vocabulary file.

• merges_file (str) – Path to the merges file.

• entity_vocab_file (str) – Path to the entity vocabulary file.

• task (str, optional) – Task for which you want to prepare sequences. One of "entity_classification", "entity_pair_classification", or "entity_span_classification". If you specify this argument, the entity sequence is automatically created based on the given entity span(s).

• max_entity_length (int, optional, defaults to 32) – The maximum length of entity_ids.

• max_mention_length (int, optional, defaults to 30) – The maximum number of tokens inside an entity span.

• entity_token_1 (str, optional, defaults to <ent>) – The special token used to represent an entity span in a word token sequence. This token is only used when task is set to "entity_classification" or "entity_pair_classification".

• entity_token_2 (str, optional, defaults to <ent2>) – The special token used to represent an entity span in a word token sequence. This token is only used when task is set to "entity_pair_classification".

__call__(text: Union[str, List[str]], text_pair: Optional[Union[str, List[str]]] = None, entity_spans: Optional[Union[List[Tuple[int, int]], List[List[Tuple[int, int]]]]] = None, entity_spans_pair: Optional[Union[List[Tuple[int, int]], List[List[Tuple[int, int]]]]] = None, entities: Optional[Union[List[str], List[List[str]]]] = None, entities_pair: Optional[Union[List[str], List[List[str]]]] = None, add_special_tokens: bool = True, padding: Union[bool, str, transformers.file_utils.PaddingStrategy] = False, truncation: Union[bool, str, transformers.tokenization_utils_base.TruncationStrategy] = False, max_length: Optional[int] = None, max_entity_length: Optional[int] = None, stride: int = 0, is_split_into_words: Optional[bool] = False, pad_to_multiple_of: Optional[int] = None, return_tensors: Optional[Union[str, transformers.file_utils.TensorType]] = None, return_token_type_ids: Optional[bool] = None, return_attention_mask: Optional[bool] = None, return_overflowing_tokens: bool = False, return_special_tokens_mask: bool = False, return_offsets_mapping: bool = False, return_length: bool = False, verbose: bool = True, **kwargs) → transformers.tokenization_utils_base.BatchEncoding[source]

Main method to tokenize and prepare for the model one or several sequence(s) or one or several pair(s) of sequences, depending on the task you want to prepare them for.

Parameters
• text (str, List[str], List[List[str]]) – The sequence or batch of sequences to be encoded. Each sequence must be a string. Note that this tokenizer does not support tokenization based on pretokenized strings.

• text_pair (str, List[str], List[List[str]]) – The sequence or batch of sequences to be encoded. Each sequence must be a string. Note that this tokenizer does not support tokenization based on pretokenized strings.

• entity_spans (List[Tuple[int, int]], List[List[Tuple[int, int]]], optional) – The sequence or batch of sequences of entity spans to be encoded. Each sequence consists of tuples each with two integers denoting character-based start and end positions of entities. If you specify "entity_classification" or "entity_pair_classification" as the task argument in the constructor, the length of each sequence must be 1 or 2, respectively. If you specify entities, the length of each sequence must be equal to the length of each sequence of entities.

• entity_spans_pair (List[Tuple[int, int]], List[List[Tuple[int, int]]], optional) – The sequence or batch of sequences of entity spans to be encoded. Each sequence consists of tuples each with two integers denoting character-based start and end positions of entities. If you specify the task argument in the constructor, this argument is ignored. If you specify entities_pair, the length of each sequence must be equal to the length of each sequence of entities_pair.

• entities (List[str], List[List[str]], optional) – The sequence or batch of sequences of entities to be encoded. Each sequence consists of strings representing entities, i.e., special entities (e.g., [MASK]) or entity titles of Wikipedia (e.g., Los Angeles). This argument is ignored if you specify the task argument in the constructor. The length of each sequence must be equal to the length of each sequence of entity_spans. If you specify entity_spans without specifying this argument, the entity sequence or the batch of entity sequences is automatically constructed by filling it with the [MASK] entity.

• entities_pair (List[str], List[List[str]], optional) – The sequence or batch of sequences of entities to be encoded. Each sequence consists of strings representing entities, i.e., special entities (e.g., [MASK]) or entity titles of Wikipedia (e.g., Los Angeles). This argument is ignored if you specify the task argument in the constructor. The length of each sequence must be equal to the length of each sequence of entity_spans_pair. If you specify entity_spans_pair without specifying this argument, the entity sequence or the batch of entity sequences is automatically constructed by filling it with the [MASK] entity.

• max_entity_length (int, optional) – The maximum length of entity_ids.

• add_special_tokens (bool, optional, defaults to True) – Whether or not to encode the sequences with the special tokens relative to their model.

• padding (bool, str or PaddingStrategy, optional, defaults to False) –

Activates and controls padding. Accepts the following values:

• True or 'longest': Pad to the longest sequence in the batch (or no padding if only a single sequence if provided).

• 'max_length': Pad to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided.

• False or 'do_not_pad' (default): No padding (i.e., can output a batch with sequences of different lengths).

• truncation (bool, str or TruncationStrategy, optional, defaults to False) –

Activates and controls truncation. Accepts the following values:

• True or 'longest_first': Truncate to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. This will truncate token by token, removing a token from the longest sequence in the pair if a pair of sequences (or a batch of pairs) is provided.

• 'only_first': Truncate to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. This will only truncate the first sequence of a pair if a pair of sequences (or a batch of pairs) is provided.

• 'only_second': Truncate to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. This will only truncate the second sequence of a pair if a pair of sequences (or a batch of pairs) is provided.

• False or 'do_not_truncate' (default): No truncation (i.e., can output batch with sequence lengths greater than the model maximum admissible input size).

• max_length (int, optional) –

Controls the maximum length to use by one of the truncation/padding parameters.

If left unset or set to None, this will use the predefined model maximum length if a maximum length is required by one of the truncation/padding parameters. If the model has no specific maximum input length (like XLNet) truncation/padding to a maximum length will be deactivated.

• stride (int, optional, defaults to 0) – If set to a number along with max_length, the overflowing tokens returned when return_overflowing_tokens=True will contain some tokens from the end of the truncated sequence returned to provide some overlap between truncated and overflowing sequences. The value of this argument defines the number of overlapping tokens.

• is_split_into_words (bool, optional, defaults to False) – Whether or not the input is already pre-tokenized (e.g., split into words). If set to True, the tokenizer assumes the input is already split into words (for instance, by splitting it on whitespace) which it will tokenize. This is useful for NER or token classification.

• pad_to_multiple_of (int, optional) – If set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).

• return_tensors (str or TensorType, optional) –

If set, will return tensors instead of list of python integers. Acceptable values are:

• 'tf': Return TensorFlow tf.constant objects.

• 'pt': Return PyTorch torch.Tensor objects.

• 'np': Return Numpy np.ndarray objects.

• return_token_type_ids (bool, optional) –

Whether to return token type IDs. If left to the default, will return the token type IDs according to the specific tokenizer’s default, defined by the return_outputs attribute.

What are token type IDs?

• return_attention_mask (bool, optional) –

Whether to return the attention mask. If left to the default, will return the attention mask according to the specific tokenizer’s default, defined by the return_outputs attribute.

• return_overflowing_tokens (bool, optional, defaults to False) – Whether or not to return overflowing token sequences.

• return_special_tokens_mask (bool, optional, defaults to False) – Whether or not to return special tokens mask information.

• return_offsets_mapping (bool, optional, defaults to False) –

Whether or not to return (char_start, char_end) for each token.

This is only available on fast tokenizers inheriting from PreTrainedTokenizerFast, if using Python’s tokenizer, this method will raise NotImplementedError.

• return_length (bool, optional, defaults to False) – Whether or not to return the lengths of the encoded inputs.

• verbose (bool, optional, defaults to True) – Whether or not to print more information and warnings.

• **kwargs – passed to the self.tokenize() method

Returns

A BatchEncoding with the following fields:

• input_ids – List of token ids to be fed to a model.

What are input IDs?

• token_type_ids – List of token type ids to be fed to a model (when return_token_type_ids=True or if “token_type_ids” is in self.model_input_names).

What are token type IDs?

• attention_mask – List of indices specifying which tokens should be attended to by the model (when return_attention_mask=True or if “attention_mask” is in self.model_input_names).

• entity_ids – List of entity ids to be fed to a model.

What are input IDs?

• entity_position_ids – List of entity positions in the input sequence to be fed to a model.

• entity_token_type_ids – List of entity token type ids to be fed to a model (when return_token_type_ids=True or if “entity_token_type_ids” is in self.model_input_names).

What are token type IDs?

• entity_attention_mask – List of indices specifying which entities should be attended to by the model (when return_attention_mask=True or if “entity_attention_mask” is in self.model_input_names).

• entity_start_positions – List of the start positions of entities in the word token sequence (when task="entity_span_classification").

• entity_end_positions – List of the end positions of entities in the word token sequence (when task="entity_span_classification").

• overflowing_tokens – List of overflowing tokens sequences (when a max_length is specified and return_overflowing_tokens=True).

• num_truncated_tokens – Number of tokens truncated (when a max_length is specified and return_overflowing_tokens=True).

• special_tokens_mask – List of 0s and 1s, with 1 specifying added special tokens and 0 specifying regular sequence tokens (when add_special_tokens=True and return_special_tokens_mask=True).

• length – The length of the inputs (when return_length=True)

Return type

BatchEncoding

save_vocabulary(save_directory: str, filename_prefix: Optional[str] = None) → Tuple[str][source]

Save only the vocabulary of the tokenizer (vocabulary + added tokens).

This method won’t save the configuration and special token mappings of the tokenizer. Use _save_pretrained() to save the whole state of the tokenizer.

Parameters
• save_directory (str) – The directory in which to save the vocabulary.

• filename_prefix (str, optional) – An optional prefix to add to the named of the saved files.

Returns

Paths to the files saved.

Return type

Tuple(str)

## LukeModel¶

class transformers.LukeModel(config, add_pooling_layer=True)[source]

The bare LUKE model transformer outputting raw hidden-states for both word tokens and entities without any specific head on top.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

Parameters

config (LukeConfig) – Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

forward(input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, entity_ids=None, entity_attention_mask=None, entity_token_type_ids=None, entity_position_ids=None, head_mask=None, inputs_embeds=None, output_attentions=None, output_hidden_states=None, return_dict=None)[source]

The LukeModel forward method, overrides the __call__() special method.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Parameters
• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) –

Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LukeTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

What are input IDs?

• attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) –

Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,

• 0 for tokens that are masked.

• token_type_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) –

Segment token indices to indicate first and second portions of the inputs. Indices are selected in [0, 1]:

• 0 corresponds to a sentence A token,

• 1 corresponds to a sentence B token.

What are token type IDs?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) –

Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• entity_ids (torch.LongTensor of shape (batch_size, entity_length)) –

Indices of entity tokens in the entity vocabulary.

Indices can be obtained using LukeTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

• entity_attention_mask (torch.FloatTensor of shape (batch_size, entity_length), optional) –

Mask to avoid performing attention on padding entity token indices. Mask values selected in [0, 1]:

• 1 for entity tokens that are not masked,

• 0 for entity tokens that are masked.

• entity_token_type_ids (torch.LongTensor of shape (batch_size, entity_length), optional) –

Segment token indices to indicate first and second portions of the entity token inputs. Indices are selected in [0, 1]:

• 0 corresponds to a portion A entity token,

• 1 corresponds to a portion B entity token.

• entity_position_ids (torch.LongTensor of shape (batch_size, entity_length, max_mention_length), optional) – Indices of positions of each input entity in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

• inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Optionally, instead of passing input_ids you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert input_ids indices into associated vectors than the model’s internal embedding lookup matrix.

• head_mask (torch.FloatTensor of shape (num_heads,) or (num_layers, num_heads), optional) –

Mask to nullify selected heads of the self-attention modules. Mask values selected in [0, 1]:

• output_attentions (bool, optional) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

• output_hidden_states (bool, optional) – Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

• return_dict (bool, optional) – Whether or not to return a ModelOutput instead of a plain tuple.

Returns

A BaseLukeModelOutputWithPooling or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LukeConfig) and inputs.

• last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) – Sequence of hidden-states at the output of the last layer of the model.

• entity_last_hidden_state (torch.FloatTensor of shape (batch_size, entity_length, hidden_size)) – Sequence of entity hidden-states at the output of the last layer of the model.

• pooler_output (torch.FloatTensor of shape (batch_size, hidden_size)) – Last layer hidden-state of the first token of the sequence (classification token) further processed by a Linear layer and a Tanh activation function.

• hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) – Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• entity_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) – Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, entity_length, hidden_size). Entity hidden-states of the model at the output of each layer plus the initial entity embedding outputs.

• attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length + entity_length, sequence_length + entity_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Examples:

>>> from transformers import LukeTokenizer, LukeModel

>>> tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base")
>>> model = LukeModel.from_pretrained("studio-ousia/luke-base")

# Compute the contextualized entity representation corresponding to the entity mention "Beyoncé"
>>> text = "Beyoncé lives in Los Angeles."
>>> entity_spans = [(0, 7)]  # character-based entity span corresponding to "Beyoncé"

>>> encoding = tokenizer(text, entity_spans=entity_spans, add_prefix_space=True, return_tensors="pt")
>>> outputs = model(**encoding)
>>> word_last_hidden_state = outputs.last_hidden_state
>>> entity_last_hidden_state = outputs.entity_last_hidden_state

# Input Wikipedia entities to obtain enriched contextualized representations of word tokens
>>> text = "Beyoncé lives in Los Angeles."
>>> entities = ["Beyoncé", "Los Angeles"]  # Wikipedia entity titles corresponding to the entity mentions "Beyoncé" and "Los Angeles"
>>> entity_spans = [(0, 7), (17, 28)]  # character-based entity spans corresponding to "Beyoncé" and "Los Angeles"

>>> encoding = tokenizer(text, entities=entities, entity_spans=entity_spans, add_prefix_space=True, return_tensors="pt")
>>> outputs = model(**encoding)
>>> word_last_hidden_state = outputs.last_hidden_state
>>> entity_last_hidden_state = outputs.entity_last_hidden_state


Return type

BaseLukeModelOutputWithPooling or tuple(torch.FloatTensor)

## LukeForEntityClassification¶

class transformers.LukeForEntityClassification(config)[source]

The LUKE model with a classification head on top (a linear layer on top of the hidden state of the first entity token) for entity classification tasks, such as Open Entity.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

Parameters

config (LukeConfig) – Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

forward(input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, entity_ids=None, entity_attention_mask=None, entity_token_type_ids=None, entity_position_ids=None, head_mask=None, inputs_embeds=None, labels=None, output_attentions=None, output_hidden_states=None, return_dict=None)[source]

The LukeForEntityClassification forward method, overrides the __call__() special method.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Parameters
• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) –

Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LukeTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

What are input IDs?

• attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) –

Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,

• 0 for tokens that are masked.

• token_type_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) –

Segment token indices to indicate first and second portions of the inputs. Indices are selected in [0, 1]:

• 0 corresponds to a sentence A token,

• 1 corresponds to a sentence B token.

What are token type IDs?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) –

Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• entity_ids (torch.LongTensor of shape (batch_size, entity_length)) –

Indices of entity tokens in the entity vocabulary.

Indices can be obtained using LukeTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

• entity_attention_mask (torch.FloatTensor of shape (batch_size, entity_length), optional) –

Mask to avoid performing attention on padding entity token indices. Mask values selected in [0, 1]:

• 1 for entity tokens that are not masked,

• 0 for entity tokens that are masked.

• entity_token_type_ids (torch.LongTensor of shape (batch_size, entity_length), optional) –

Segment token indices to indicate first and second portions of the entity token inputs. Indices are selected in [0, 1]:

• 0 corresponds to a portion A entity token,

• 1 corresponds to a portion B entity token.

• entity_position_ids (torch.LongTensor of shape (batch_size, entity_length, max_mention_length), optional) – Indices of positions of each input entity in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

• inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Optionally, instead of passing input_ids you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert input_ids indices into associated vectors than the model’s internal embedding lookup matrix.

• head_mask (torch.FloatTensor of shape (num_heads,) or (num_layers, num_heads), optional) –

Mask to nullify selected heads of the self-attention modules. Mask values selected in [0, 1]:

• output_attentions (bool, optional) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

• output_hidden_states (bool, optional) – Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

• return_dict (bool, optional) – Whether or not to return a ModelOutput instead of a plain tuple.

• labels (torch.LongTensor of shape (batch_size,) or (batch_size, num_labels), optional) – Labels for computing the classification loss. If the shape is (batch_size,), the cross entropy loss is used for the single-label classification. In this case, labels should contain the indices that should be in [0, ..., config.num_labels - 1]. If the shape is (batch_size, num_labels), the binary cross entropy loss is used for the multi-label classification. In this case, labels should only contain [0, 1], where 0 and 1 indicate false and true, respectively.

Returns

A EntityClassificationOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LukeConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) – Classification loss.

• logits (torch.FloatTensor of shape (batch_size, config.num_labels)) – Classification scores (before SoftMax).

• hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) – Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• entity_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) – Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, entity_length, hidden_size). Entity hidden-states of the model at the output of each layer plus the initial entity embedding outputs.

• attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Examples:

>>> from transformers import LukeTokenizer, LukeForEntityClassification

>>> tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-large-finetuned-open-entity")
>>> model = LukeForEntityClassification.from_pretrained("studio-ousia/luke-large-finetuned-open-entity")

>>> text = "Beyoncé lives in Los Angeles."
>>> entity_spans = [(0, 7)]  # character-based entity span corresponding to "Beyoncé"
>>> inputs = tokenizer(text, entity_spans=entity_spans, return_tensors="pt")
>>> outputs = model(**inputs)
>>> logits = outputs.logits
>>> predicted_class_idx = logits.argmax(-1).item()
>>> print("Predicted class:", model.config.id2label[predicted_class_idx])
Predicted class: person


Return type

EntityClassificationOutput or tuple(torch.FloatTensor)

## LukeForEntityPairClassification¶

class transformers.LukeForEntityPairClassification(config)[source]

The LUKE model with a classification head on top (a linear layer on top of the hidden states of the two entity tokens) for entity pair classification tasks, such as TACRED.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

Parameters

config (LukeConfig) – Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

forward(input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, entity_ids=None, entity_attention_mask=None, entity_token_type_ids=None, entity_position_ids=None, head_mask=None, inputs_embeds=None, labels=None, output_attentions=None, output_hidden_states=None, return_dict=None)[source]

The LukeForEntityPairClassification forward method, overrides the __call__() special method.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Parameters
• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) –

Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LukeTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

What are input IDs?

• attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) –

Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,

• 0 for tokens that are masked.

• token_type_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) –

Segment token indices to indicate first and second portions of the inputs. Indices are selected in [0, 1]:

• 0 corresponds to a sentence A token,

• 1 corresponds to a sentence B token.

What are token type IDs?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) –

Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• entity_ids (torch.LongTensor of shape (batch_size, entity_length)) –

Indices of entity tokens in the entity vocabulary.

Indices can be obtained using LukeTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

• entity_attention_mask (torch.FloatTensor of shape (batch_size, entity_length), optional) –

Mask to avoid performing attention on padding entity token indices. Mask values selected in [0, 1]:

• 1 for entity tokens that are not masked,

• 0 for entity tokens that are masked.

• entity_token_type_ids (torch.LongTensor of shape (batch_size, entity_length), optional) –

Segment token indices to indicate first and second portions of the entity token inputs. Indices are selected in [0, 1]:

• 0 corresponds to a portion A entity token,

• 1 corresponds to a portion B entity token.

• entity_position_ids (torch.LongTensor of shape (batch_size, entity_length, max_mention_length), optional) – Indices of positions of each input entity in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

• inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Optionally, instead of passing input_ids you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert input_ids indices into associated vectors than the model’s internal embedding lookup matrix.

• head_mask (torch.FloatTensor of shape (num_heads,) or (num_layers, num_heads), optional) –

Mask to nullify selected heads of the self-attention modules. Mask values selected in [0, 1]:

• output_attentions (bool, optional) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

• output_hidden_states (bool, optional) – Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

• return_dict (bool, optional) – Whether or not to return a ModelOutput instead of a plain tuple.

• labels (torch.LongTensor of shape (batch_size,) or (batch_size, num_labels), optional) – Labels for computing the classification loss. If the shape is (batch_size,), the cross entropy loss is used for the single-label classification. In this case, labels should contain the indices that should be in [0, ..., config.num_labels - 1]. If the shape is (batch_size, num_labels), the binary cross entropy loss is used for the multi-label classification. In this case, labels should only contain [0, 1], where 0 and 1 indicate false and true, respectively.

Returns

A EntityPairClassificationOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LukeConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) – Classification loss.

• logits (torch.FloatTensor of shape (batch_size, config.num_labels)) – Classification scores (before SoftMax).

• hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) – Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• entity_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) – Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, entity_length, hidden_size). Entity hidden-states of the model at the output of each layer plus the initial entity embedding outputs.

• attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Examples:

>>> from transformers import LukeTokenizer, LukeForEntityPairClassification

>>> tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-large-finetuned-tacred")
>>> model = LukeForEntityPairClassification.from_pretrained("studio-ousia/luke-large-finetuned-tacred")

>>> text = "Beyoncé lives in Los Angeles."
>>> entity_spans = [(0, 7), (17, 28)]  # character-based entity spans corresponding to "Beyoncé" and "Los Angeles"
>>> inputs = tokenizer(text, entity_spans=entity_spans, return_tensors="pt")
>>> outputs = model(**inputs)
>>> logits = outputs.logits
>>> predicted_class_idx = logits.argmax(-1).item()
>>> print("Predicted class:", model.config.id2label[predicted_class_idx])
Predicted class: per:cities_of_residence


Return type

EntityPairClassificationOutput or tuple(torch.FloatTensor)

## LukeForEntitySpanClassification¶

class transformers.LukeForEntitySpanClassification(config)[source]

The LUKE model with a span classification head on top (a linear layer on top of the hidden states output) for tasks such as named entity recognition.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

Parameters

config (LukeConfig) – Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

forward(input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, entity_ids=None, entity_attention_mask=None, entity_token_type_ids=None, entity_position_ids=None, entity_start_positions=None, entity_end_positions=None, head_mask=None, inputs_embeds=None, labels=None, output_attentions=None, output_hidden_states=None, return_dict=None)[source]

The LukeForEntitySpanClassification forward method, overrides the __call__() special method.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Parameters
• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) –

Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LukeTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

What are input IDs?

• attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) –

Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,

• 0 for tokens that are masked.

• token_type_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) –

Segment token indices to indicate first and second portions of the inputs. Indices are selected in [0, 1]:

• 0 corresponds to a sentence A token,

• 1 corresponds to a sentence B token.

What are token type IDs?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) –

Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• entity_ids (torch.LongTensor of shape (batch_size, entity_length)) –

Indices of entity tokens in the entity vocabulary.

Indices can be obtained using LukeTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

• entity_attention_mask (torch.FloatTensor of shape (batch_size, entity_length), optional) –

Mask to avoid performing attention on padding entity token indices. Mask values selected in [0, 1]:

• 1 for entity tokens that are not masked,

• 0 for entity tokens that are masked.

• entity_token_type_ids (torch.LongTensor of shape (batch_size, entity_length), optional) –

Segment token indices to indicate first and second portions of the entity token inputs. Indices are selected in [0, 1]:

• 0 corresponds to a portion A entity token,

• 1 corresponds to a portion B entity token.

• entity_position_ids (torch.LongTensor of shape (batch_size, entity_length, max_mention_length), optional) – Indices of positions of each input entity in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

• inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Optionally, instead of passing input_ids you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert input_ids indices into associated vectors than the model’s internal embedding lookup matrix.

• head_mask (torch.FloatTensor of shape (num_heads,) or (num_layers, num_heads), optional) –

Mask to nullify selected heads of the self-attention modules. Mask values selected in [0, 1]:

• output_attentions (bool, optional) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

• output_hidden_states (bool, optional) – Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

• return_dict (bool, optional) – Whether or not to return a ModelOutput instead of a plain tuple.

• entity_start_positions (torch.LongTensor) – The start positions of entities in the word token sequence.

• entity_end_positions (torch.LongTensor) – The end positions of entities in the word token sequence.

• labels (torch.LongTensor of shape (batch_size, entity_length) or (batch_size, entity_length, num_labels), optional) – Labels for computing the classification loss. If the shape is (batch_size, entity_length), the cross entropy loss is used for the single-label classification. In this case, labels should contain the indices that should be in [0, ..., config.num_labels - 1]. If the shape is (batch_size, entity_length, num_labels), the binary cross entropy loss is used for the multi-label classification. In this case, labels should only contain [0, 1], where 0 and 1 indicate false and true, respectively.

Returns

A EntitySpanClassificationOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LukeConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) – Classification loss.

• logits (torch.FloatTensor of shape (batch_size, config.num_labels)) – Classification scores (before SoftMax).

• hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) – Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• entity_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) – Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, entity_length, hidden_size). Entity hidden-states of the model at the output of each layer plus the initial entity embedding outputs.

• attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Examples:

>>> from transformers import LukeTokenizer, LukeForEntitySpanClassification

>>> tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-large-finetuned-conll-2003")
>>> model = LukeForEntitySpanClassification.from_pretrained("studio-ousia/luke-large-finetuned-conll-2003")

>>> text = "Beyoncé lives in Los Angeles"

# List all possible entity spans in the text
>>> word_start_positions = [0, 8, 14, 17, 21]  # character-based start positions of word tokens
>>> word_end_positions = [7, 13, 16, 20, 28]  # character-based end positions of word tokens
>>> entity_spans = []
>>> for i, start_pos in enumerate(word_start_positions):
...     for end_pos in word_end_positions[i:]:
...         entity_spans.append((start_pos, end_pos))

>>> inputs = tokenizer(text, entity_spans=entity_spans, return_tensors="pt")
>>> outputs = model(**inputs)
>>> logits = outputs.logits
>>> predicted_class_indices = logits.argmax(-1).squeeze().tolist()
>>> for span, predicted_class_idx in zip(entity_spans, predicted_class_indices):
...     if predicted_class_idx != 0:
...        print(text[span[0]:span[1]], model.config.id2label[predicted_class_idx])
Beyoncé PER
Los Angeles LOC


Return type

EntitySpanClassificationOutput or tuple(torch.FloatTensor)