Granular Boundary Encoding for Sequence Labeling

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 216 papers21 conceptsfrozen 2026-10-04

In natural language processing, many tasks boil down to identifying spans of text—whether they are named entities, clinical concepts, or individual words in a continuous stream of characters. The standard way to teach a model to recognize these spans is through a tagging scheme. While the simple BIO (Begin, Inside, Outside) format is a common baseline, a growing body of research suggests that increasing the granularity of these tags significantly improves a model's ability to locate boundaries.

The Ambiguity of Simple Tagging

The BIO scheme is intuitive: the first token of an entity is marked B, subsequent tokens are I, and everything else is O. However, this simplicity creates a structural blind spot. In BIO, the model is never explicitly told when an entity ends; it only knows an entity has finished because it encounters an O or a new B.

This lack of an explicit "end" marker makes it difficult for models to distinguish between the internal structure of a long entity and the transition between two adjacent entities of the same type. As Dan Roth and Lev Ratinov point out in Design Challenges and Misconceptions in Named Entity Recognition, using a more expressive representation like BILOU (Begin, Inside, Last, Outside, Unit) provides critical structural information that helps the classifier make more precise decisions.

Building Intuition with BILOU and BIOES

To solve the boundary problem, researchers have converged on schemes that add specific tags for the end of a span and for single-token entities. The two most prominent are BILOU and BIOES (Begin, Inside, Outside, End, Single).

  • The End Marker (L/E): By marking the final token of a multi-word phrase, the model learns the specific patterns that signal a conclusion. Hongxin Cao and colleagues demonstrated in Recognizing clinical entities in hospital discharge summaries that this reduces errors in identifying where complex medical phrases terminate.
  • The Single-Token Marker (U/S): In BIO, a single-word entity is just a B. In BILOU, it is a U (Unit). This distinction is vital because the linguistic features surrounding a standalone entity (like "London") are often very different from those surrounding the start of a long phrase (like "The University of...").

Kurt Junshean Espinosa and Sophia Ananiadou found that this granularity is especially helpful for noisy data, such as tweets, where explicit boundary cues help the model navigate informal syntax Learning to recognise named entities in tweets.

Internal Structure in Word Segmentation

The benefits of granular tagging extend beyond word-level entities to character-level tasks like Chinese Word Segmentation (CWS). Because Chinese does not use spaces, a model must decide which characters cluster into words.

Standard systems might use four tags (Begin, Middle, End, Single). However, Hai Zhao and Changning Huang showed in An Improved Chinese Word Segmentation System with Conditional Random Field that expanding to a six-tag set—adding specific markers for the second and third positions in a word—allows the model to capture the internal structural patterns of long compound terms. This is echoed by Xiao Qin and Xiaojun Wan, who found that increased granularity is particularly beneficial for technical or scientific literature where long, complex terms are frequent CRF-based Experiments for Cross-Domain Chinese Word Segmentation.

Why It Works: Informative Transitions

From a machine learning perspective, these extra tags work because they constrain the search space of the model. In a Conditional Random Field (CRF) or a BiLSTM-CRF, the model learns transition probabilities between tags.

In a simple BIO scheme, a B can be followed by an I, an O, or another B. In a BILOU scheme, the rules are stricter: a B must eventually be followed by an L. As Namrata Nath and Ivan Lee argue in their work on clinical concept extraction, this increased specificity helps the model distinguish between adjacent or nested concepts in dense, unstructured text NEAR: Named entity and attribute recognition of clinical concepts.

By providing the model with a richer "vocabulary" of labels, researchers like Rodrigo Agerri and German Rigau have shown that even general-purpose systems can be adapted to highly specialized domains, such as medical disability annotation, with minimal hand-engineered rules Simple Language Independent Sequence Labelling. The tags themselves act as a form of structural supervision, guiding the model to respect the natural boundaries of language.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 18 papers
  2. Version 1 · 2026-10-04 · 16 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_granular_boundary_tagging_7181bf87,
  title = {granular boundary tagging},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/granular-boundary-tagging/rdr_7181bf87c19a462bb752e5520e19d877},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_granular_boundary_encoding_for_sequence__7181bf87_v1,
  title = {Granular Boundary Encoding for Sequence Labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/granular-boundary-encoding-for-sequence-labeling/rdr_7181bf87c19a462bb752e5520e19d877/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}