TL;DR
This paper explores how introductory Natural Language Processing (NLP) courses can effectively balance classical NLP techniques with modern deep learning approaches. By comparing two university courses, the authors argue that teaching foundational, classical methods is crucial for building student intuition and a solid understanding of NLP problems, even as deep learning dominates the field.
What Problem Does This Paper Address?
The rapid advancement of deep learning, particularly Transformer-based models, has revolutionized NLP. This has led to a question among educators: what is the role of traditional, "classical" NLP approaches (those predating deep learning) in modern university curricula? Students themselves sometimes question the relevance of learning older methods when state-of-the-art results are achieved with deep learning. This paper addresses this by investigating how introductory NLP courses can strike a balance between these two paradigms, ensuring students gain both practical deep learning skills and a robust conceptual foundation.
Methods
The authors employ a comparative case study approach, analyzing two introductory NLP courses: one at the University of New South Wales (UNSW) in Australia and another at the Indian Institute of Technology Bombay (IITB) in India. They examine:
- Lecture Plans: How classical and deep learning topics are interleaved or sequenced.
- Tutorials: The role of practical sessions in reinforcing concepts.
- Programming Assessments: How assignments are designed to test understanding of both classical and deep learning techniques.
- Course Projects: The structure and evaluation of student projects, encouraging the use of diverse methods.
They also draw parallels to the "objects-first vs. objects-later" debate in computer science education to inform their discussion on pedagogical strategies.
Datasets
Not discussed in the paper. The paper focuses on pedagogical approaches rather than specific model training or evaluation on datasets.
Key Insights and Ideas
- Classical Approaches Build Intuition: Teaching classical methods (like rule-based systems, statistical models, HMMs, CRFs) helps students grasp the fundamental challenges of NLP tasks, understand the "why" behind solutions, and appreciate the value addition of deep learning.
- Cognitive Load Theory is Relevant: Similar to introductory programming, introducing complex concepts gradually (as classical methods can offer) can reduce cognitive overload and improve learning assimilation.
- Balance is Key, Not Exclusion: While deep learning is state-of-the-art, excluding classical approaches entirely can hinder a deep understanding. The paper advocates for an interleaved or carefully sequenced approach.
- Context Matters: The optimal balance can depend on factors like course length, student background, and whether it's the sole NLP offering or part of a broader sequence.
- Student Motivation: Classical approaches can serve as a powerful motivator by highlighting the limitations of older methods, thus underscoring the revolutionary impact of deep learning.
- Annotation as a Bridge: Classical techniques remain relevant for annotation tasks, providing a linguistic rigor that helps students understand the nuances of NLP data.
Key Results
The paper presents qualitative findings from the comparison of the two courses:
- NLP-UNSW: Interleaves deep learning with classical approaches from week 1, introducing NLP tasks via black-box libraries first, then delving into representations, Transformer architecture, and finally task-specific applications with classical and deep learning methods. The group project encourages experimentation with simpler models alongside deep learning.
- NLP-IITB: Starts with classical approaches (HMMs, probabilistic parsing) before moving to machine translation and Transformers. The individual assignments cover both statistical and deep learning models, and the course project emphasizes selecting the "right" algorithm.
- Textbook Analysis: Most NLP textbooks cover classical approaches, often focusing on statistical models, before introducing deep learning.
- Course Outline Analysis: Publicly available course outlines show varying degrees of emphasis on classical topics, with some explicitly covering them as foundational.
Limitations and Future Work
- Limited Sample Size: The paper analyzes only two courses and a subset of NLP textbooks, acknowledging that these are examples and not necessarily prescriptive for all educators.
- Generalizability: The findings are derived from specific course contexts and may not directly apply to all institutions or student demographics.
- Future Work: The authors hope their analysis will help NLP educators find the right balance between classical and deep learning approaches in their curricula. They suggest that further research into pedagogical strategies for NLP is warranted.
Summary updated