TL;DR
This paper introduces a practical, hands-on curriculum for teaching Natural Language Processing (NLP) and Information Retrieval (IR) to biomedical informatics (BMI) students. By using a sequence of Jupyter notebooks, the curriculum balances theoretical understanding with practical application, preparing students to leverage NLP tools in their research.
What Problem Does This Paper Address?
Biomedical informatics (BMI) research increasingly relies on Natural Language Processing (NLP) for tasks like extracting information from scientific literature or clinical notes. However, introducing BMI students to NLP presents a challenge: how to balance the deep computational and linguistic understanding of NLP with the practical, application-focused needs of BMI researchers? Many BMI students are not computer science experts and may not have prior AI or NLP background. This paper addresses this gap by proposing a structured, accessible curriculum that equips these students with the fundamental NLP and IR skills they need for their domain.
Methods
The authors developed a set of three Jupyter notebooks designed as a single assignment for an introductory BMI course. The notebooks guide students through core NLP and IR concepts and workflows:
- Notebook 1: Fundamentals of document analysis: Introduces basic NLP preprocessing (tokenization, stemming, casing, stop-word removal) and two indexing techniques: inverted indexing and TF-IDF (Term Frequency-Inverse Document Frequency). Students then implement a synthetic IR task and evaluate system performance using standard metrics like recall and precision.
- Notebook 2: Introduction to word embeddings: Explains word embeddings as a text representation method. Students learn to create embeddings using Singular Value Decomposition (SVD) on a corpus and visualize them to understand semantic similarity. They also explore pre-trained word2vec embeddings.
- Notebook 3: Introduction to BERT and clinicalBERT: Introduces students to transformer-based language models like BERT and clinicalBERT. Using YouTube tutorials for background on neural networks, students work with the Hugging Face Transformers library to implement Named Entity Recognition and Medical Language Inference tasks, comparing BERT and clinicalBERT on a clinical dataset.
The approach emphasizes using pre-built libraries (like NLTK and Transformers) to allow students to focus on application rather than building models from scratch.
Datasets
- Natural Language Toolkit (NLTK) corpus: Used in Notebook 1 for basic preprocessing and indexing tasks.
- Synthetic IR dataset: A collection of 12 documents and 20 queries created for evaluation in Notebook 1.
- MedNLI dataset: Created from MIMIC-III clinical notes, used in Notebook 3 for the Medical Language Inference task.
Key Insights and Ideas
- Bridging the gap: The curriculum successfully balances theoretical NLP/IR concepts with practical application, making it accessible to students with diverse backgrounds.
- Modular workflow: The activities demonstrate a generalizable NLP workflow (preprocessing, indexing, execution, evaluation) applicable beyond IR.
- Importance of domain-specific models: The introduction to clinicalBERT highlights the value of pre-training models on domain-specific data for improved performance in specialized fields like medicine.
- Hands-on learning: Using Jupyter notebooks with code samples and guided questions fosters active learning and deeper understanding.
- Ethical considerations: The authors acknowledge the growing importance of discussing ethical implications of AI in medicine, a crucial area for future refinement.
Key Results
The paper doesn't present quantitative results in terms of model performance metrics. Instead, the key "results" are the successful implementation of a pedagogical approach that:
- Exposes introductory BMI students to fundamental text representation and language model strategies for IR.
- Provides hands-on experience in creating NLP workflows using pre-built tools.
- Successfully guides students through tasks like document analysis, word embedding visualization, and applying BERT for clinical NLP tasks.
Limitations and Future Work
- Limited dataset availability: The authors note the challenge of using real-world clinical datasets due to Data Use Agreements and training requirements. They used the NLTK corpus for initial notebooks and MIMIC-III for the BERT notebook, acknowledging its limitations.
- Ethical considerations: The authors plan to investigate and incorporate discussions on the ethical implications of biomedical AI technologies, including data privacy and responsible use of NLP tools.
- Further refinement: The authors are continuously refining and extending these teaching materials.
- Broader applicability: While focused on IR, the modular workflow is designed to be applicable to other NLP applications in BMI.
The materials are available at https://github.com/dbmi-pitt/bioinf_teachingNLP.
Summary updated