# A boundary enhanced multi-task neural attention approach for Chinese named entity recognition

## TL;DR

This paper introduces a new multi-task neural attention approach to improve Chinese named entity recognition by jointly detecting word boundaries and identifying entities.

## Contribution

The novel contribution is combining boundary detection with NER in a joint framework using hybrid embeddings and a feature fusion layer.

## Key findings

- The proposed method outperforms existing models on Chinese NER tasks.
- Joint training with boundary prediction improves entity recognition accuracy.
- Hybrid embeddings and attention mechanisms enhance contextual understanding.

## Abstract

Named Entity Recognition (NER) stands as a fundamental task in Chinese information processing. However, it encounters unique difficulties due to the lack of explicit word boundaries in the Chinese language. This study proposes framing Chinese NER as a joint task that combines boundary detection and entity identification within an encoder-decoder architecture. The presented method utilizes hybrid embeddings to enhance word-level representations and naturally incorporates head and tail boundary information to improve NER performance. It combines two types of tasks: sequence labeling for NER and binary classification for boundary prediction. In the primary NER task, a convolutional attention network serves as the encoder to extract contextual information about the target word from the input. For the auxiliary boundary prediction task, two Bi-GRU networks are employed to model long range semantic associations and predict the start and end of entities. A feature fusion layer is then introduced to adjust the contribution of the main and auxiliary tasks to the hidden states of the global representation. The final input representation, obtained through the joint training framework where the learned boundary information supports the NER task, is passed to the CRF decoding layer. Experimental results on the Weibo and Ontonotes5.0 datasets show that the multi - task learning framework significantly enhances Chinese NER performance compared to existing models.