[Paper Review] A taxonomy and review of generalization research in NLP

Dieuwke Hupkes, Mario Giulianelli+18 more|arXiv (Cornell University)|Oct 6, 2022

Natural Language Processing Techniques629 references41 citations

TL;DR

This paper introduces a comprehensive five-axis taxonomy to systematize generalization research in NLP, analyzing over 400 papers and 600 experiments. It identifies key motivations, types of generalization, data shifts, sources of shift, and loci within the modeling pipeline, revealing critical gaps in current evaluation practices and advocating for standardized, robust generalization testing as the new norm in NLP research.

ABSTRACT

The ability to generalise well is one of the primary desiderata of natural language processing (NLP). Yet, what 'good generalisation' entails and how it should be evaluated is not well understood, nor are there any evaluation standards for generalisation. In this paper, we lay the groundwork to address both of these issues. We present a taxonomy for characterising and understanding generalisation research in NLP. Our taxonomy is based on an extensive literature review of generalisation research, and contains five axes along which studies can differ: their main motivation, the type of generalisation they investigate, the type of data shift they consider, the source of this data shift, and the locus of the shift within the modelling pipeline. We use our taxonomy to classify over 400 papers that test generalisation, for a total of more than 600 individual experiments. Considering the results of this review, we present an in-depth analysis that maps out the current state of generalisation research in NLP, and we make recommendations for which areas might deserve attention in the future. Along with this paper, we release a webpage where the results of our review can be dynamically explored, and which we intend to update as new NLP generalisation studies are published. With this work, we aim to take steps towards making state-of-the-art generalisation testing the new status quo in NLP.

Motivation & Objective

• To address the lack of consensus and standardization in evaluating generalization in NLP models.
• To identify and categorize the diverse motivations behind generalization research, including robustness, fairness, and cognitive plausibility.
• To develop a systematic taxonomy that captures the full spectrum of generalization types, data shifts, and their sources and loci in the modeling pipeline.
• To analyze the current state of generalization research by classifying over 400 papers and 600 experiments using the proposed taxonomy.
• To recommend future research directions and promote the adoption of systematic generalization evaluation as the new standard in NLP.

Proposed method

• The authors developed a five-axis taxonomy to characterize generalization research: (1) motivation, (2) type of generalization, (3) data shift type, (4) source of shift, and (5) locus of shift in the modeling pipeline.
• The taxonomy was applied to systematically classify 400+ NLP papers and over 600 individual experiments from the literature.
• The classification was based on a comprehensive literature review spanning multiple NLP subfields and evaluation paradigms.
• The authors analyzed patterns in the data to identify imbalances in research focus, such as over-reliance on i.i.d. splits and underrepresentation of compositional and structural generalization.
• They released an interactive, dynamically updatable web platform to visualize and explore the taxonomy and results, supporting ongoing research curation.
• The method integrates qualitative categorization with quantitative analysis of research trends to inform future evaluation standards.

Experimental results

Research questions

• RQ1 What are the dominant motivations driving generalization research in NLP, and how do they shape evaluation design?
• RQ2 What types of generalization—e.g., compositional, structural, cross-lingual—are most commonly studied, and which remain underexplored?
• RQ3 How do data shifts in NLP generalization studies differ in terms of source (natural vs. generated) and type (covariate, label, full shift)?
• RQ4 Where in the modeling pipeline—pre-training, training, or test—do data shifts most commonly occur, and how does this affect model evaluation?
• RQ5 Which generalization capabilities are currently under-investigated despite their importance for robust, human-like model behavior?

Key findings

• Over 60% of generalization studies focus on simple data shifts like domain or style changes, while compositional and structural generalization remain significantly under-researched.
• Only 12% of studies evaluate models on fully generated data splits (e.g., SCAN), despite their value in probing compositional generalization.
• The majority of studies (78%) evaluate generalization using natural data splits or naturally occurring shifts, but these often lack systematic control and reproducibility.
• There is a strong imbalance in research focus: robustness and fairness are prominent motivations, while cognitive and intrinsic motivations (e.g., human-like reasoning) are underrepresented.
• Models achieve high performance on standard i.i.d. benchmarks (e.g., GLUE scores >90), yet fail on non-i.i.d. or distributionally shifted data, indicating a critical gap between performance and true generalization.
• The taxonomy reveals that 85% of generalization studies involve shifts from training to test data, with only 10% examining shifts from pre-training to training or pre-training to test, highlighting a blind spot in evaluating lifelong or continual learning capabilities.