gsri-18/medical-ner-bleurt-separated

Medical NER Dataset with BLEURT-Based Quality Separation

Dataset Summary

This dataset is a processed version of the gsri-18/body_parts_ner_dataset_synthetic dataset. The original dataset contains synthetic sentences with BIO-labeled body parts for Named Entity Recognition (NER) tasks.

This version extends the original by adding a BLEURT score to each sample. The score measures the semantic coherence of each sentence against a reference sentence representing a normal, healthy state. The dataset is then labeled into two quality classes based on this score, making it useful for data filtering, quality analysis, or training models to identify high-quality medical text.

The creation process is documented in the provided Jupyter Notebook.

How the Dataset Was Created

The dataset was generated using the following process:

1. Load Source Data: The initial dataset gsri-18/body_parts_ner_dataset_synthetic was loaded.
2. BLEURT Scoring: Each sentence was scored using the bleurt-base-512 model. The score was calculated based on the semantic similarity to the following reference sentence:

   "The patient reported normal function and good health without any complications."

3. Class Separation: Based on the BLEURT score, each sample was assigned to one of two classes:
   • positive_bleurt (Class 1): Samples with a BLEURT score greater than or equal to 0. These sentences are considered more coherent or semantically similar to the reference.
   • negative_bleurt (Class 0): Samples with a BLEURT score less than 0. These sentences are considered less coherent or semantically dissimilar.
4. New Fields Added: Three new columns were added to the dataset to store this information: bleurt_score, bleurt_class, and bleurt_class_name.

Data Splits

The dataset is divided into the same splits as the original:

Split	Number of Samples
train	108,000
test	8,001
dev	8,001

Dataset Structure

Each sample in the dataset has the following fields:

• sentence_id: A unique integer identifier for the sentence.
• sentence: The full text of the sentence.
• tokens: A list of strings representing the tokenized sentence.
• bio_labels: A list of strings with the BIO (Beginning, Inside, Outside) tags for each token (e.g., 'B-BodyPart', 'I-BodyPart', 'O').
• int_labels: Integer representations of the bio_labels (0: 'O', 1: 'B-BodyPart', 2: 'I-BodyPart').
• bleurt_score: A float value representing the calculated BLEURT score.
• bleurt_class: An integer label (1 for positive, 0 for negative) based on the BLEURT score.
• bleurt_class_name: A string ('positive_bleurt' or 'negative_bleurt') corresponding to the class.

Data Sample

Here is an example from the training set:

{
  "sentence_id": 1,
  "sentence": "During the checkup, the patient complained of discomfort in his Peripheral nerves, with slight swelling around the Temporal lobe and Cerebellum.",
  "tokens": ["During", "the", "checkup", ",", "the", "patient", "complained", "of", "discomfort", "in", "his", "Peripheral", "nerves", ",", "with", "slight", "swelling", "around", "the", "Temporal", "lobe", "and", "Cerebellum", "."],
  "bio_labels": ["O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-BodyPart", "I-BodyPart", "O", "O", "O", "O", "O", "O", "B-BodyPart", "I-BodyPart", "O", "B-BodyPart", "O"],
  "int_labels": [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0, 0, 1, 2, 0, 1, 0],
  "bleurt_score": -0.82470703125,
  "bleurt_class": 0,
  "bleurt_class_name": "negative_bleurt"
}

How to Use the Dataset

You can load and use this dataset with the datasets library as follows:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("gsri-18/medical-ner-bleurt-separated")

# Access a split
train_data = dataset['train']

# Print the first example
print(train_data[0])

# You can easily filter the dataset by BLEURT class
positive_samples = train_data.filter(lambda example: example['bleurt_class'] == 1)
negative_samples = train_data.filter(lambda example: example['bleurt_class'] == 0)

print(f"Number of positive samples in train set: {len(positive_samples)}")
print(f"Number of negative samples in train set: {len(negative_samples)}")