Fine‑Grained Entity Recognition in Tencent TexSmart: System Overview and Key Techniques

System Overview TexSmart is a Chinese‑English natural language understanding platform comparable to LTP, FastNLP, and CoreNLP, offering lexical, syntactic, and semantic analysis. It distinguishes itself with three main features: (1) fine‑grained NER supporting thousands of entity types and hierarchical relations, (2) enhanced semantic understanding that provides context‑aware associations, and (3) design for both academic precision and industrial speed requirements.

Fine‑Grained NER The system can identify not only coarse entities such as “南昌” (city) and “流浪地球” (movie) but also composite entities and their sub‑components, addressing scalability and ambiguity challenges by expanding the type inventory to over a thousand categories.

Knowledge‑Base Combination Method An unsupervised approach extracts "is‑a" pairs from raw text, maps them to predefined type hierarchies, and uses similarity between mentions and type candidates for disambiguation. A hybrid strategy combines a small supervised model for coarse categories with unsupervised fine‑grained prediction, yielding consistent performance gains.

Remote Supervision via Similar Entities The method treats similar mentions as "Sibling Mentions" and models the problem as a heterogeneous graph with Mention and Type nodes. Graph neural networks with self‑attention and dropout on edges learn node representations for classification, improving robustness on ambiguous or short contexts.

Zero‑Shot Multi‑Source Fusion To handle unseen entity types, the approach integrates three auxiliary information sources: context‑consistency (BERT), type hierarchy (hierarchy‑aware transformer encoder), and background knowledge (prototypes, descriptions via WordNet). These modules are combined through a multi‑premise textual entailment framework and a Trans‑style loss, achieving 4–5 point gains on BBN and Wiki benchmarks.

Experiments & Results Extensive ablation studies demonstrate the contribution of each auxiliary source, the benefit of longer contexts, and the effectiveness of the hybrid graph method. The system consistently outperforms baselines on both supervised and unsupervised settings.

Q&A Highlights The authors discussed future work on span design, the origin of prototypes, and alternative GNN architectures for heterogeneous graphs.

References [1] TexSmart official site, [2] LTP, [3] FastNLP, [4] Stanford CoreNLP, [5] Entity typing literature, etc.