The central challenge in supervised learning is the "annotation bottleneck": the high cost and slow speed of manually labeling large datasets. A recurring solution across diverse domains—from biomedical indexing to sarcasm detection—is to treat classification not as a direct mapping from features to labels, but as a retrieval and propagation task. The core intuition is simple: if a new, unlabeled instance is semantically or structurally similar to a known, labeled instance, it likely shares the same category, tag, or attribute.
The Retrieval-First Paradigm
Instead of training a global parametric model to predict labels from scratch, this approach uses a reference set of human-annotated examples as a searchable database. When a new document or image arrives, the system identifies its $k$-nearest neighbors in this reference set and "propagates" their labels to the new instance.
This mechanism is particularly effective in specialized domains like biomedical literature. For example, Huang et al. demonstrated that aggregating tags from a document's nearest neighbors in a pre-annotated corpus significantly outperforms standard benchmarks for Medical Subject Headings (MeSH). Similarly, Zhu et al. found that search-based labeling—retrieving similar entries and borrowing their metadata—is more effective for automated indexing than direct concept extraction. By leveraging existing human-curated metadata, systems can capture nuanced labeling patterns that simple keyword matching might miss, as noted by Kamineni et al..
Defining Similarity Across Modalities
The success of label propagation depends on how "similarity" is defined. Researchers have employed various metrics to bridge the gap between raw data and semantic labels:
- String and Structural Similarity: In spoken dialog systems, Sarikaya et al. used string-similarity metrics to transfer semantic tags between sentences. For opinion mining, Xu et al. utilized edit distance and word coverage to identify structurally similar sentences and transfer opinion holders and targets.
- Embedding Spaces: Modern approaches often project data into a vector space. Swarup et al. used sentence embeddings to retrieve labeled instances for semantic parsing in scientific procedural text. In narrative classification, Younus et al. leveraged story embeddings to transfer narrative frames between semantically coherent news articles.
- Visual Vocabularies: Even in non-textual domains, the principle holds. Hare et al. treated visual features as words, using information retrieval techniques to propagate keywords from visually similar images to unlabeled queries.
Distributional and Relational Propagation
Beyond direct similarity, labels can be propagated through the principle of substitutability or co-occurrence. Haghighi et al. proposed distributional label propagation for grammar induction, where labels are assigned to unknown spans based on their contextual similarity to known "prototypes." If two phrases appear in identical contexts, they likely belong to the same syntactic category.
Similarly, Zhuang et al. used relational label propagation to elicit affective events from language models. By identifying events that co-occur in generated text, the system can assign attributes to new data points based on the known properties of their "seed" neighbors.
Efficiency and Human-in-the-Loop Bootstrapping
Label propagation is not just a classification strategy; it is a powerful tool for label-efficient annotation. The LEAN-LIFE framework by Lee et al. uses semantic similarity to locate "trigger words" provided by annotators, automatically suggesting labels for similar contexts across massive unlabeled datasets.
To handle the computational cost of searching millions of records, researchers have introduced high-speed retrieval mechanisms. Mao et al. used similarity-preserved hashing to make topic labeling nearly instantaneous. More recently, Tian et al. argued that mimicking human indexing through similar record matching (MaRSI) is not only more interpretable but also significantly faster than fine-tuning large language models for extreme multi-label tasks.
By reframing classification as a retrieval problem, these methods bypass the need for exhaustive manual training, instead borrowing the "wisdom" already embedded in existing datasets.
Key Foundational Work
Fast semi-automatic semantic annotation for spoken dialog systems (2004) — contributed 2 concept elements to this direction.
Title generation for spoken broadcast news using a training corpus (2000) — contributed 2 concept elements to this direction.