The Fragmentation of Label Spaces
Machine learning models are frequently confined to "isolated islands" of data (Li et al.). Even when multiple datasets capture the exact same underlying real-world phenomena, they rarely share a common vocabulary. For instance, one emotion recognition dataset might label a text as "furious," another might use "anger," and a third might simply classify it as "negative."
This taxonomy fragmentation creates artificial domain gaps. When models are trained on a single dataset, they often fail to generalize because they memorize dataset-specific annotation quirks and surface forms rather than learning universal concepts (Augenstein et al.). To build truly robust models, researchers use cross-dataset taxonomy mapping—a suite of techniques designed to harmonize disparate label spaces into a unified, standardized representation.
Coarsening and Hierarchical Alignment
The most direct way to bridge inconsistent taxonomies is to map fine-grained, disparate labels into broader, shared functional super-classes. This structural alignment creates a consistent output space that allows a model to learn shared features across otherwise incompatible data sources.
For example, in dialogue act classification, specific conversational labels can be mapped to high-level intent categories to facilitate cross-domain evaluation (Webb). Similarly, in emotion prediction, consolidating fragmented, fine-grained categories into a core framework (such as joy, anger, fear, and sadness) allows for the aggregation of diverse social media corpora into a single, cohesive training set (Bianchi et al.).
To preserve semantic nuance without sacrificing unification, researchers often arrange labels into graph structures or lexical hierarchies. By mapping diverse labels to common intermediate nodes within a hierarchy, systems can resolve naming conflicts and synonyms across heterogeneous databases (Bear et al.; Tommasi). This structural mapping ensures that a model can treat semantically equivalent objects as a single class, dramatically increasing the effective volume of training data.
Projecting to Continuous and Latent Spaces
Categorical mapping can sometimes be too rigid, particularly when datasets use entirely different annotation philosophies or granularities. When discrete alignment fails, an alternative mechanism is to project disparate labels onto a unified continuous scale or coordinate system.
In speech emotion recognition, for example, mapping categorical labels onto a dimensional coordinate system allows models to apply distance-based loss functions across completely different data sources (Gao). A similar strategy works for factuality prediction, where discrete annotations from multiple datasets are projected onto a single continuous scale to enable cross-corpus evaluation (Stanovsky et al.).
When explicit mapping functions are unavailable, unsupervised label alignment can be achieved through embedding spaces. By calculating the similarity between pre-trained label embeddings, disparate or noisy labeling schemes can be mapped to a standardized taxonomy automatically. This technique allows high-quality, specialized models to benefit from the broader patterns found in larger, less structured corpora without requiring manual re-annotation.
Data-Driven and Automatic Unification
Rather than relying entirely on manual alignment or expert-crafted dictionaries, modern frameworks can discover taxonomy mappings automatically by observing how models behave on actual data.
A prominent example of this is found in object detection. The authors of Simple Multi-dataset Detection demonstrated that taxonomy unification can be achieved by analyzing the visual correlations between predictions from different dataset-specific classifiers. If two distinct classifiers consistently fire on the same visual objects across images, the system automatically identifies those labels as candidates for merging into a single unified concept. This data-driven alignment bypasses the need for manual label merging while allowing a single model to train on multiple massive, independently annotated datasets simultaneously.
Why Taxonomy Mapping Works
Taxonomy mapping works because it fundamentally alters what the model is forced to learn. When a model is trained on a single dataset with a closed taxonomy, it can achieve high accuracy by exploiting spurious correlations or memorizing specific annotation biases.
By forcing the model to operate within a unified label space across heterogeneous datasets, the training objective shifts from learning dataset-specific annotation boundaries to learning invariant structural features. Whether adapting across different hospital documentation styles (Zhou et al.) or cross-platform developer interests (Lee et al.), taxonomy harmonization acts as a powerful regularizer. It strips away the superficial noise of how data was labeled, leaving behind the core semantic patterns that generalize across domains.
Key Foundational Work
Investigating the Portability of Corpus-Derived Cue Phrases for Dialogue Act Classification (2008) — contributed 2 concept elements to this direction.