The final project of EMNLP course in PKU, 2021 Spring.
Entity and relation extraction aims to identify named entities from plain texts and relations among them. These two task are typically modeled jointly with a neural network in an end-to-end manner. Recent work ? models the two tasks separately, and propose a pipeline where the entity information is fused to relation extractor at the input layer. Surprisingly, that model outperforms a lot of strong joint models. In this work, we follow their separate modeling style, and propose an entity-specific encoding with multi-head attention mechanism. The model (BERT-JRE) reaches results comparable with input-level fusion in ? on relation extraction, but are much more time-efficient.
Please refer to the report.
To run the code, please check run.sh and install required packages before you run the script.