gbpatentdata/patent_titles_ner

Patent Title Extraction Model

Model Description

patent_titles_ner is a fine-tuned XLM-RoBERTa-large model that has been trained on a custom dataset of OCR'd front pages of patent specifications published by the British Patent Office, and filed between 1617-1899. It has been trained to recognize the stated titles of inventions.

We take the original xlm-roberta-large weights and fine tune on our custom dataset for 15 epochs with a learning rate of 6e-05 and a batch size of 21. We chose the learning rate by tuning on the validation set.

Usage

This model can be used with HuggingFace Transformer's Pipelines API for NER:

from transformers import pipeline, AutoModelForTokenClassification, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("gbpatentdata/patent_titles_ner")
model = AutoModelForTokenClassification.from_pretrained("gbpatentdata/patent_titles_ner")


def custom_recognizer(text, model=model, tokenizer=tokenizer, device=0):

    # HF ner pipeline
    token_level_results = pipeline("ner", model=model, device=0, tokenizer=tokenizer)(text)

    # keep entities tracked
    entities = []
    current_entity = None

    for item in token_level_results:

        tag = item['entity']

        # replace '▁' with space for easier reading (_ is created by the XLM-RoBERTa tokenizer)
        word = item['word'].replace('▁', ' ')

        # aggregate I-O-B tagged entities
        if tag.startswith('B-'):

            if current_entity:
                entities.append(current_entity)

            current_entity = {'type': tag[2:], 'text': word.strip(), 'start': item['start'], 'end': item['end']}

        elif tag.startswith('I-'):

            if current_entity and tag[2:] == current_entity['type']:
                current_entity['text'] += word
                current_entity['end'] = item['end']

            else:

                if current_entity:
                    entities.append(current_entity)

                current_entity = {'type': tag[2:], 'text': word.strip(), 'start': item['start'], 'end': item['end']}

        else:
            # deal with O tag
            if current_entity:
                entities.append(current_entity)
            current_entity = None

    if current_entity:
        # add to entities
        entities.append(current_entity)

    # track entity merges
    merged_entities = []

    # merge entities of the same type
    for entity in entities:
        if merged_entities and merged_entities[-1]['type'] == entity['type'] and merged_entities[-1]['end'] == entity['start']:
            merged_entities[-1]['text'] += entity['text']
            merged_entities[-1]['end'] = entity['end']
        else:
            merged_entities.append(entity)

    # clean up extra spaces
    for entity in merged_entities:
        entity['text'] = ' '.join(entity['text'].split())

    # convert to list of dicts
    return [{'class': entity['type'],
             'entity_text': entity['text'],
             'start': entity['start'],
             'end': entity['end']} for entity in merged_entities]



example = """
Date of Application, 1st Aug., 1890-Accepted, 6th Sept., 1890
COMPLETE SPECIFICATION.
Improvements in Coin-freed Apparatus for the Sale of Goods.
I, CHARLES LOTINGA, of 33 Cambridge Street, Lower Grange, Cardiff, in the County of Glamorgan, Gentleman,
do hereby declare the nature of this invention and in what manner the same is to be performed,
to be particularly described and ascertained in and by the following statement
"""

ner_results = custom_recognizer(example)
print(ner_results)

Training Data

The custom dataset of front page texts of patent specifications was assembled in the following steps:

1. We fine tuned a YOLO vision model to detect bounding boxes around text. We use this to identify text regions on the front pages of patent specifications.
2. We use Google Cloud Vision to OCR the detected text regions, and then concatenate the OCR text.
3. We randomly sample 200 front page texts (and another 201 oversampled from those that contain either firm or communicant information).

Our custom dataset has accurate manual labels generated by a graduate student. The final dataset is split 60-20-20 (train-val-test). In the event that the front page text is too long, we restrict the text to the first 512 tokens.

Evaluation

Our evaluation metric is F1 at the full entity-level. That is, we aggregated adjacent-indexed entities into full entities and computed F1 scores requiring an exact match. These scores for the test set are below.

Full Entity	Precision	Recall	F1-Score
TITLE	93.9%	97.5%	95.7%