Tibert

^home ^projects

Github

Tibert is a transformers-compatible reproduction from the paper End-to-end Neural Coreference Resolution (1) with several modifications. Among these:

It can be installed with pip install tibert.

Here is an example of using the simple prediction interface: 

from tibert import BertForCoreferenceResolution, predict_coref_simple
from tibert.utils import pprint_coreference_document
from transformers import BertTokenizerFast

model = BertForCoreferenceResolution.from_pretrained(
    "compnet-renard/bert-base-cased-literary-coref"
)
tokenizer = BertTokenizerFast.from_pretrained("bert-base-cased")

annotated_doc = predict_coref_simple(
    "Sli did not want the earpods. He didn't like them.", model, tokenizer
)

pprint_coreference_document(annotated_doc)

results in: 

>>> (0 Sli ) did not want the earpods. (0 He ) didn't like them.

A more advanced prediction interface is available: 

from transformers import BertTokenizerFast
from tibert import predict_coref, BertForCoreferenceResolution

model = BertForCoreferenceResolution.from_pretrained(
    "compnet-renard/bert-base-cased-literary-coref"
)
tokenizer = BertTokenizerFast.from_pretrained("bert-base-cased")

documents = [
    "Sli did not want the earpods. He didn't like them.",
    "Princess Liana felt sad, because Zarth Arn was gone. The princess went to sleep.",
]

annotated_docs = predict_coref(documents, model, tokenizer, batch_size=2)

for doc in annotated_docs:
    pprint_coreference_document(doc)

results in: 

>>> (0 Sli ) did not want the earpods . (0 He ) didn't like them .
>>> (0 Princess Liana ) felt sad , because (1 Zarth Arn ) was gone . (0 The princess) went to sleep .

The coreference chains predicted can be accessed using the .coref_chains attribute: 

annotated_doc = predict_coref_simple(
    "Princess Liana felt sad, because Zarth Arn was gone. The princess went to sleep.",
    model,
    tokenizer
)
print(annotated_doc.coref_chains)

>>>[[Mention(tokens=['The', 'princess'], start_idx=11, end_idx=13), Mention(tokens=['Princess', 'Liana'], start_idx=0, end_idx=2)], [Mention(tokens=['Zarth', 'Arn'], start_idx=6, end_idx=8)]]

Hierarchical Merging

Hierarchical merging allows to reduce RAM usage and computations when performing inference on long documents. To do so, the user provides the text cut in chunks. The model will perform prediction for chunks, which means the long document wont be taken at once into memory. Then, hierarchical merging will try to merge chunk predictions. This allow scaling to arbitrarily large documents. See Gupta et al. 2024 (4) for more details. Hierarchical merging can be used as follows: 

from tibert import BertForCoreferenceResolution, predict_coref
from tibert.utils import pprint_coreference_document
from transformers import BertTokenizerFast

model = BertForCoreferenceResolution.from_pretrained(
    "compnet-renard/bert-base-cased-literary-coref"
)
tokenizer = BertTokenizerFast.from_pretrained("bert-base-cased")

chunk1 = "Princess Liana felt sad, because Zarth Arn was gone."
chunk2 = "She went to sleep."

annotated_doc = predict_coref(
    [chunk1, chunk2], model, tokenizer, hierarchical_merging=True
)

pprint_coreference_document(annotated_doc)

This results in: 

>>>(1 Princess Liana ) felt sad , because (0 Zarth Arn ) was gone . (1 She ) went to sleep .

Even if the mentions Princess Liana and She are not in the same chunk, hierarchical merging still resolves this case correctly.

Training a model

Aside from the tibert.train.train_coref_model function, it is possible to train a model from the command line. Training a model requires installing the sacred library. Here is the most basic example: 

python -m tibert.run_train with\
       dataset_path=/path/to/litbank/repository\
       out_model_dir=/path/to/output/model/directory

The following parameters can be set (taken from ./tibert/run_train.py config function):

Parameter Default Value
batch_size 1
epochs_nb 30
dataset_name "litbank"
dataset_path "~/litbank"
mentions_per_tokens 0.4
antecedents_nb 350
max_span_size 10
mention_scorer_hidden_size 3000
sents_per_documents_train 11
mention_loss_coeff 0.1
bert_lr 1e-5
task_lr 2e-4
dropout 0.3
segment_size 128
encoder "bert-base-cased"
out_model_dir "~/tibert/model"
checkpoint None

One can monitor training metrics by adding run observers using command line flags - see sacred documentation for more details.

Method

We reimplemented the model from Lee et al. 2017 (1) from scratch, but used BERT as the encoder as in Joshi et al. 2019 (2). We do not use higher order inference as in Lee et al. 2018 (5) since it was found to be not necessarily useful by Xu and Choi 2020 (6).

Singletons

Unfortunately, the framework from Lee et al. 2017 (1) cannot represent singletons. This is because the authors were working on the OntoNotes dataset (7), where singletons are not annotated. We wanted to work on Litbank (8), so we had to find a way to represent singletons.

We opted to do as in Xu and Choi 2021 (3): we consider mention with a high enough mention scores as singletons, even when they are in no clusters. To force the model to learn proper mention scores, we add an auxiliary loss on mention score (as in Xu and Choi 2021 (3)). To counter dataset imbalance between positive and negative mentions, we opt to compute a weighted loss instead of performing sampling.

Additional Features

Several work make use of additional features. For now, only the distance between spans is implemented.

Results

The following table presents the results we obtained on Litbank by training this model. We evaluate on 10% of Litbank documents, each of which consists of ~2000 tokens. The split column indicate whether documents were split in blocks of 512 tokens. The HM coumns indicates whether we use hierarchical merging.

Dataset Base model split HM MUC B3 CEAF BLANC LEA time (m:s)
Litbank bert-base-cased no no 75.03 60.66 48.71 62.96 32.84 22:07
Litbank bert-base-cased yes no 73.84 49.14 47.88 48.41 27.63 16:18
Litbank bert-base-cased yes yes 74.54 59.30 46.98 62.69 42.46 21:13

References