dell-research-harvard/historical_newspaper_ner

Historical newspaper NER

Model description

historical_newspaper_ner is a fine-tuned Roberta-large model for use on text that may contain OCR errors.

It has been trained to recognize four types of entities: location (LOC), organizations (ORG), person (PER) and Miscellaneous (MISC).

It was trained on a custom historical newspaper dataset, with highly accurate labels. All data were double entered by two highly skilled Harvard undergraduates and all discrepancies were resolved by hand.

Intended uses

You can use this model with Transformers pipeline for NER.

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

tokenizer = AutoTokenizer.from_pretrained("dell-research-harvard/historical_newspaper_ner")
model = AutoModelForTokenClassification.from_pretrained("dell-research-harvard/historical_newspaper_ner")

nlp = pipeline("ner", model=model, tokenizer=tokenizer)
example = "My name is Wolfgang and I live in Berlin"

ner_results = nlp(example)
print(ner_results)

Limitations and bias

This model was trained on historical news and may reflect biases from a specific period of time. It may also not generalise well to other setting. Additionally, the model occasionally tags subword tokens as entities and post-processing of results may be necessary to handle those cases.

Training data

The training dataset distinguishes between the beginning and continuation of an entity so that if there are back-to-back entities of the same type, the model can output where the second entity begins. Each token will be classified as one of the following classes:

Abbreviation	Description
O	Outside of a named entity
B-MISC	Beginning of a miscellaneous entity
I-MISC	Miscellaneous entity
B-PER	Beginning of a person’s name
I-PER	Person’s name
B-ORG	Beginning of an organization
I-ORG	organization
B-LOC	Beginning of a location
I-LOC	Location

This model was fine-tuned on historical English-language news that had been OCRd from American newspapers. Unlike other NER datasets, this data has highly accurate labels. All data were double entered by two highly skilled Harvard undergraduates and all discrepancies were resolved by hand.

# of training examples per entity type

Dataset	Article	PER	ORG	LOC	MISC
Train	227	1345	450	1191	1037
Dev	48	231	59	192	149
Test	48	261	83	199	181

Training procedure

The data was used to fine-tune a Roberta-Large model (Liu et. al, 2020) at a learning rate of 4.7e-05 with a batch size of 128 for 184 epochs.

Eval results

entities	f1
PER	94.3
ORG	80.7
LOC	90.8
MISC	79.6
Overall (stringent)	86.5
Overall (ignoring entity type)	90.4

Notes

This model card was influence by that of dslim/bert-base-NER

Citation

If you use this model, you can cite the following paper:

@misc{franklin2024ndjv,
      title={News Deja Vu: Connecting Past and Present with Semantic Search}, 
      author={Brevin Franklin, Emily Silcock, Abhishek Arora, Tom Bryan and Melissa Dell},
      year={2024},
      eprint={2406.15593},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.15593}, 
}

Applications

We applied this model to a century of historical news articles. You can see all the named entities in the NEWSWIRE dataset.