🧑‍🔬 Jianlin Su

Jianlin Su works on neural methods for natural language processing, with papers here on two connected problems. One is how transformer models represent the order of tokens in a sequence. The other is how to pull structured facts out of plain sentences. Both reflect an interest in rethinking the core design of a model rather than adding parts on top of it, and both rely on careful mathematical framing of what the model has to capture.

The most cited of these papers is RoFormer: Enhanced Transformer with Rotary Position Embedding (2021). It introduced Rotary Position Embedding, or RoPE. Instead of adding a position vector to each token, RoPE rotates query and key vectors by an angle tied to absolute position, so that their inner product reflects only the relative distance between tokens. The idea is derived with complex numbers in two dimensions and extended by pairing coordinates. The method also suits linear self-attention and sequences of varying length. The earlier Cascade Binary Tagging Framework for Relational Triple Extraction (2020) takes up a problem in information extraction: recovering subject, relation and object triples when several facts in one sentence share entities. Earlier joint models classified relations for each entity pair, and the framework moves away from that approach because of its class imbalance and its confusion over overlapping triples.

The newest work here is the rotary embedding paper, which turns from extraction tasks toward the internal mechanics of transformers. Su's bylines most recently list Southeast University, after earlier links with Moonshot AI and with Sun Yat-sen University. Co-authors on these papers include Yunfeng Liu, Bo Wen and Yue Wang.

Jianlin Su has 2 papers in our library, in 2020–2021, most of them on Information extraction; often with Bo Wen, Yuan Tian, Yue Wang.