🧑‍🔬 Xiaonan Li

Xiaonan Li works in natural language processing. Their papers here, from 2020 to 2023, deal with named entity recognition for Chinese text and with how language models can be helped by better choices of input. Both lines of work draw on Transformer-based architectures and on ideas from information retrieval. A recurring concern is replacing complicated or task-specific machinery with simpler, more general designs that run efficiently.

The most cited of these papers is FLAT: Chinese NER Using Flat-Lattice Transformer. Chinese has no spaces between words, so entity recognition there often relies on lattices that combine characters with words matched from a lexicon. Earlier lattice models were recurrent or structurally intricate and made poor use of GPU parallelism. FLAT turns the lattice into a flat collection of spans, each marked by where it begins and ends. A relative position encoding built from these boundaries lets a Transformer capture how the spans relate to one another.

The later work moves from sequence labelling to in-context learning, where a language model solves a task from a few examples placed in its prompt. Unified Demonstration Retriever for In-Context Learning (2023) addresses the fact that results depend strongly on which examples are picked. Instead of generic similarity search or a separately trained retriever for every task, it trains a single bi-encoder shared across many tasks. Each example is prefixed with a short task instruction, and training signals from different tasks are brought into one common form.

The bylines of Li's papers include Fudan University among their affiliations. Co-authors on the site include Xipeng Qiu, Hang Yan and Xuanjing Huang.

Xiaonan Li has 2 papers in our library, in 2020–2023, most of them on Named entity recognition, Language models & pretraining, Information retrieval & search; often with Hang Yan, Xipeng Qiu, Xuanjing Huang.