metadata

tags:
  - feature-extraction
  - endpoints-template
license: apache-2.0
library_name: generic

Coreference Resolution for Long Documents

Modified coreference resolution model from BERT for Coreference Resolution: Baselines and Analysis for handling long documents (~40K words) efficiently (500K words/s on a NVIDIA Tesla V100). The checkpoint is based on AllenNLP's coref-spanbert-large-2021.03.10. This modified model was used in DAPR: A Benchmark on Document-Aware Passage Retrieval.

Usage

API call

One can call the Hugging's Inference Endpoints API directly: (need your access token from https://huggingface.co/settings/tokens and the loading takes around 6 minutes)

import requests
import time

API_URL = "https://api-inference.huggingface.co/models/kwang2049/long-coref"
headers = {"Authorization": "Bearer ${YOUR_HUGGINGFACE_ACCESS_TOKEN}"}


def query(payload):
    while True:
        response = requests.post(API_URL, headers=headers, json=payload)
        if response.status_code == 503:
            time.sleep(5)
            print(response.json()["error"])
            continue
        elif response.status_code == 200:
            return response.json()
        else:
            error_message = f"{response.status_code}: {response.json['error']}."
            raise requests.HTTPError(error_message)


doc = [
    "The Half Moon is a public house and music venue in Putney, London. It is one of the city's longest running live music venues, and has hosted live music every night since 1963.",
    "The pub is on the south side of the Lower Richmond road, in the London Borough of Wandsworth."
]

PARAGRAPH_DELIMITER = "\n\n"

output = query(
    {
        "inputs": PARAGRAPH_DELIMITER.join(doc),
    }
)
print(output)
# {
#     'pargraph_sentences': ..., 
#     'top_spans': ..., 
#     'antecedents': ...
# }

Local run

One can also run the code of the repo on a local machine:

# Clone the repo
git lfs install
git clone https://huggingface.co/kwang2049/long-coref
cd long-coref
pip install -r requirements.txt
python local_run.py

Evalution

The evaluation results on OntoNotesv5 are:

Model	Precision	Recall	F1	Input Length
AllenNLP's original implementation	79.2	78.4	78.8	<= 2K words
This modification	78.9	67.0	72.4	<= 40K words

Citation

If you use the repo, feel free to cite our publication DAPR: A Benchmark on Document-Aware Passage Retrieval:

@article{wang2023dapr,
    title = "DAPR: A Benchmark on Document-Aware Passage Retrieval",
    author = "Kexin Wang and Nils Reimers and Iryna Gurevych", 
    journal= "arXiv preprint arXiv:2305.13915",
    year = "2023",
    url = "https://arxiv.org/abs/2305.13915",
}