RoBERTa zero-shot interpretability paper

Data format

The model supports standard tab seperated CoNLL format - one word per line, with tokens and labels being separated with a tab. There is a new line between sentences.

For example (CoNLL2010 uncertainty detection):

We  C
believe U
that  C
this  C
server  C
is  C
useful  C

Run

To run seq class training, please run: python seq_class_script.py {config_path} {gpu_id}

Similarly, to generate token level predictions run: python model_to_token_preds.py {config_path} {gpu_id} with two optional arguments: {layer_id} and {head_id} if running for RoBERTa's internal attention.

To evaluate on the token level, run: python token_preds_evaluate.py {config_path} {layer_id (optional)} {head_id (optional)}

Config

Sample configs are provided in the configs/ folder for all 3 stages.

Possible entries of the configs are:

experiment_name - name of the experiment being run now

dataset - name of the dataset being used

model_name - name of the HuggingFace BERT model to use

max_seq_length - maximum seq length used by the HuggingFace library

per_device_train_batch_size - train batch size

per_device_eval_batch_size - eval batch size

num_train_epochs - number of epochs

warmup_ratio - warmup ratio of the optimiser

learning_rate - training learning rate

weight_decay - weight decay inside BERT during fine tuning

seed - random seed

adam_epsilon - epsilon value for adam optimiser

test_label_dummy - dummy label to use for test dataset

make_all_labels_equal_max - make all tokens labels equal to max

is_seq_class - perform seq class or token class

lowercase - make all tokens lowercase

gradient_accumulation_steps - over how many steps to accumulate gradient

save_steps - how often to save checkpoints

logging_steps - how often to log

output_dir - where to save the model

do_mask_words - mask out some words with a token

mask_prob - probability of a word being masked

hid_to_attn_dropout - dropout between soft attention layer and BERT

attention_evidence_size - size of the attention weights layer

final_hidden_layer_size - size of the final output hidden layer

initializer_name - how to initialise weights and bias

attention_activation - function to use for soft attention activation

soft_attention - apply soft attention architecture on top

soft_attention_gamma - how much imporatance to give to optimise max token label to be equal to max sentence label

soft_attention_alpha- importance of optimising min token label to 0

square_attention - square attentions during normalisation

freeze_bert_layers_up_to - freeze all BERT layers up to x

zero_n - make at least n token labels be close to 0.0

zero_delta - importance of the zero_n loss

dataset_split - which dataset to use for predictions/eval

model_path - path to the pretrained model

datetime - datetime to use

results_input_dir - input dir of the the token-level preds

results_input_filename - file in whcih token-level preds are stored in the tsv format

importance_threshold - threshold to use for eval

top_count - only mark x top scores as positive

preds_output_filename - save predictions to file

eval_results_filanamd - save eval results to file

Acknowledgements

Please note that some of this code is based on the token prediction tutorial for BERT using HuggingFace: https://github.com/huggingface/transformers/tree/master/examples/token-classification

and Marek Rei's mltagger: https://github.com/marekrei/mltagger