Reproduce “FLAT: Chinese NER Using Flat-Lattice Transformer” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaonan Li, Hang Yan, Xipeng Qiu, Xuanjing Huang
Fudan University researchers developed FLAT, a Flat-Lattice Transformer, that efficiently incorporates lexical information into Chinese Named Entity Recognition by converting dynamic lattice structures into a flat sequence of spans with a novel relative position encoding. This approach achieved state-of-the-art F1 scores across multiple Chinese NER datasets while demonstrating superior inference speed due to improved parallelization.
Key insights
- Converting complex lattice structures into a "flat" span representation with specialized positional encodings allows for full Transformer parallelization while retaining critical lexical information.
- A specially designed relative position encoding effectively captures nuanced relationships (e.g., overlap, containment, distance) between word and character spans within the flat-lattice.
- Explicitly incorporating both content similarity and detailed relative positional relationships into the Transformer's self-attention mechanism enhances modeling of long-distance dependencies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.