BAT - Biomedical Augmentation for Text

Keywords: Data augmentation, Neuro-Symbolic AI, NLP, LLM, UMLS

---

A Toolkit for Biomedical Text Augmentation

Package Overview

This Python package consists of a Neuro-Symbolic pipeline, blending knowledge-driven and data-driven approaches.

Pipeline Components

Knowledge-Based perturbation (knowledge-driven):

• Med-synonym replacement: Replaces medical terms with one of their formalized synonyms from structured domain knowledge (UMLS Metathesaurus).
• General synonym replacement: Replaces terms with one of their general-purpose synonyms from Wordnet.

Transformer-Based perturbation (data-driven):

• Back-translation: Translates text into an intermediate language and then back into the original language using multilingual MT models.
• Contextual word prediction: Fills in masked single-token words within the input text based on the in-context predictions from BERT-based language models.
• Rephrasing: Rewrites text using the capabilities of LLMs.

Requirements

1. Unified Medical Language System® (UMLS®) License:

• Mandatory for using the Med-synonym replacement component.
• Optional for the General synonym replacement.

2. LLM Functional Block:

• A functional block with any preferred (open source or proprietary) LLM must be configured to use the Rephrasing component.
• Alternatively, you can use the default gpt-4o-mini model by providing your personal API key.

Installation

1. Make sure you have the latest version of pip installed

   pip install --upgrade pip

2. Install BiomedicalAugmentation-for-Text through pip

   pip install --index-url https://test.pypi.org/simple/ --no-deps BiomedicalAugmentation-for-Text

Usage

Here is a minimal example of how the BAT package can be invoked with BiomedicalAugmentation-for-Text.

1. Through the AugmentedSample class: A compact and streamlined interface that integrates all components into a cohesive workflow.

from bioTextAugPackage.init import *
import bioTextAugPackage.augmented_sample as aug_sample

config = Config()
input_text = "No lytic lesions are observed at the vertebral levels included in the scans. No signs of listhesis."
augmented_sample = aug_sample.AugmentedSample(config_params=config, technique_tag="TB-back_translation",
                                              src_data=input_text, src_lang="english", n_synth_data=5)
ans = augmented_sample.run()

2. By invoking individual functions: Provides more control and flexibility to apply specific components independently.

from bioTextAugPackage.init import *
import bioTextAugPackage.transformer_based_functions as tb
import  bioTextAugPackage.metrics as metrics

config = Config()
input_text = "No lytic lesions are observed at the vertebral levels included in the scans. No signs of listhesis."

src_lang = "en"
trg_lang = "fr"

mt_model_name1 = f"Helsinki-NLP/opus-mt-{src_lang}-{trg_lang}"
mt_model1 = AutoModelForSeq2SeqLM.from_pretrained(mt_model_name1)
mt_tokenizer1 = AutoTokenizer.from_pretrained(mt_model_name1)

mt_model_name2 = f"Helsinki-NLP/opus-mt-{trg_lang}-{src_lang}"
mt_model2 = AutoModelForSeq2SeqLM.from_pretrained(mt_model_name2)
mt_tokenizer2 = AutoTokenizer.from_pretrained(mt_model_name2, clean_up_tokenization_spaces=True)

ans = tb.back_translation(src_text=input_text,
                          model1=mt_model1, tokenizer1=mt_tokenizer1,
                          model2=mt_model2, tokenizer2=mt_tokenizer2)

print(ans)
overlap_score = metrics.compute_overlap(synthetic_data=ans[0], src_data=input_text, tokenizer=config.base_tokenizer)
similarity_score = metrics.compute_similarity(synthetic_data=ans[0], src_data=input_text, se_model_name=config.se_model_name)
print(f"overlap_score: {overlap_score} - similarity_score: {similarity_score}")

A more extensive example, including advanced usage, can be found in this notebook.

Contacts and Useful Links

• Repository maintainer: Laura Bergomi
  

• Project Link: https://github.com/bmi-labmedinfo/BAT

• Package Link: https://test.pypi.org/project/BiomedicalAugmentation-for-Text/

License

Distributed under MIT License. See LICENSE