The Million-Label NER: Breaking Scale Barriers with GLiNER bi-encoder
Ihor Stepanov, Mykhailo Shtopko, Dmytro Vodianytskyi, Oleksandr Lukashov · 2026
Read The Million-Label NER: Breaking Scale Barriers with GLiNER bi-encoder by Ihor Stepanov, Mykhailo Shtopko, Dmytro Vodianytskyi, Oleksandr Lukashov (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-18487.txt.
Description
Ihor Stepanov and three co-authors scale named entity recognition to massive label sets. Their 2026 paper presents GLiNER-bi-Encoder, a novel architecture joining zero-shot flexibility with industrial-scale efficiency. The original GLiNER framework generalizes strongly, but its joint-encoding approach grows quadratically complex as labels increase.
The bi-encoder design splits the work into a label encoder and a context encoder, removing the context-window bottleneck. It enables recognition of thousands of entity types at once, potentially millions, with minimal overhead. Experiments show state-of-the-art zero-shot performance, reaching sixty-one point five percent Micro-F1 on the CrossNER benchmark. With pre-computed label embeddings, throughput rises up to one hundred thirty times at one thousand twenty-four labels over uni-encoder predecessors. The team also introduces GLiNKER, a modular framework applying this architecture to high-performance entity linking across huge knowledge bases such as Wikidata.
Catalog record
| CREATOR | Ihor Stepanov, Mykhailo Shtopko, Dmytro Vodianytskyi, Oleksandr Lukashov |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-18487