Traditional Named Entity Recognition (NER) is typically framed as a sequence labeling problem. Models assign a tag (like B-PER or I-ORG) to every single token in a sentence. While this approach is standard, it forces a model to solve two fundamentally different problems simultaneously: localization (where does the entity start and end?) and classification (what kind of entity is it?).
A growing body of research suggests that "decoupling" these two concerns leads to more robust, efficient, and flexible systems. By treating NER as a two-stage pipeline—first identifying candidate spans and then labeling them—researchers are overcoming the limitations of traditional sequence labeling, particularly in data-scarce and complex domains.
The Problem with Conflation
In a standard IOB (Inside-Outside-Beginning) tagging scheme, the "Outside" (O) class usually accounts for over 90% of the data. This creates a massive class imbalance that can drown out the signal for specific entity types. Furthermore, as Jinyuan Fang and Xiaobin Wang argue in their work on the MANNER framework, conflating these tasks makes it difficult for models to adapt to new domains. If you want to recognize a new entity type, a traditional model must relearn both the boundaries and the semantics from scratch.
By contrast, the "decomposed" framework treats "entity-ness" as a structural property that is often independent of the specific category.
Stage 1: Class-Agnostic Localization
The first stage of these frameworks focuses exclusively on identifying potential entity boundaries. This is often called "span detection" or "mention detection."
The intuition here is that the linguistic markers of an entity—capitalization, syntactic position, or surrounding prepositions—are often consistent regardless of whether the entity is a "Protein," a "City," or a "Software Product." Yaqing Wang and Haoda Chu demonstrate in SpanNER that a model trained to find spans in a class-agnostic way becomes a powerful "entity seeker" that generalizes across domains.
Different mechanisms are used for this localization:
- Binary Pointers: Xiaoya Li uses a dual-pointer mechanism to predict the start and end probabilities of spans, a method that naturally handles nested entities where one span exists inside another Locate and Label.
- Boundary Matrices: Chengyu Wang and Jianing Wang utilize a global boundary matrix in SpanProto to extract all possible candidate spans before any typing occurs.
Stage 2: Semantic Classification
Once the candidate spans are "located," the second stage "labels" them. Because the search space has been narrowed from the entire document to just a few candidate spans, the model can afford to use more computationally intensive or semantically rich methods for classification.
In Split-NER, Jatin Arora and Youngja Park use a Question Answering (QA) framework for this stage. Instead of searching the whole text for "Person" entities, they simply ask the model: "What is the type of [identified span]?" This reduces redundant computations and significantly speeds up inference.
This separation also allows for the integration of external knowledge. Pratyay Banerjee and Chitta Baral show that by decoupling localization, the classification stage can be guided by external knowledge prompts or dictionary matching without confusing the model's ability to find the entity's physical boundaries Knowledge Guided NER.
Why Decomposition Works
The success of this research direction stems from three primary advantages:
- Mitigating Class Imbalance: By removing the "O" (non-entity) tokens from the classification phase, the model can focus its parameters on the nuances between similar categories (e.g., distinguishing a "Drug" from a "Chemical Compound") rather than just trying to distinguish "something" from "nothing" MANNER.
- Few-Shot and Zero-Shot Transfer: In low-resource settings, a model can be pre-trained to find spans on a large, general dataset. When moving to a specialized domain (like biochemistry), it only needs to learn the new semantic labels, not the basic structure of what an entity looks like Coach.
- Handling Complexity: For tasks like Nested NER or Coreference Resolution, the number of possible spans is $O(n^2)$. Decomposing the task allows models to filter out the vast majority of "junk" spans early, making the subsequent logic much simpler. For instance, Vladimir Dobrovolskii shows that separating coreference logic from span detection reduces computational complexity from $O(n^4)$ to $O(n^2)$ Word-Level Coreference Resolution.
By breaking the "monolithic" sequence labeling approach into specialized stages, researchers like Hanjun Luo and Xinfeng Li have demonstrated that even lightweight, open-source models can outperform massive commercial LLMs on complex, fine-grained extraction tasks DynamicNER. The "divide and conquer" strategy effectively turns a single hard problem into two manageable ones.