Related Experiment Video
Updated: Jan 1, 2026
05:38
Interaction between Phonological and Semantic Processes in Visual Word Recognition using Electrophysiology
Published on: June 29, 2021
2.8K
Improving clinical named entity recognition in Chinese using the graphical and phonetic feature
Yifei Wang1, Sophia Ananiadou2, Jun'ichi Tsujii2,3
1National Centre of Text Mining, University of Manchester, Manchester, UK. yifei.wang@manchester.ac.uk.
BMC Medical Informatics and Decision Making
|December 24, 2019
Summary
This study enhances Chinese Clinical Named Entity Recognition by incorporating graphical and phonetic features of characters. This approach improves accuracy in identifying medical terms within Chinese text.
Area of Science:
- Natural Language Processing
- Computational Linguistics
- Bioinformatics
Background:
- Clinical Named Entity Recognition (CNER) identifies medical terms in text.
- Chinese language presents unique challenges for CNER due to its character-based nature.
- Traditional methods struggle to leverage the rich graphical and phonetic information within Chinese characters.
Purpose of the Study:
- To improve Chinese Clinical Named Entity Recognition.
- To explore the utility of graphical and phonetic features for CNER.
- To evaluate embedding models incorporating these features.
Main Methods:
- Developed three distinct embedding models for Chinese CNER.
- Utilized both graphical (e.g., radicals) and phonetic (e.g., pinyin) features.
- Tested models on annotated data, varying the proportion of phono-semantic characters.
Main Results:
- The model combining primary radicals and pinyin achieved an F-measure of 0.712.
- Incorporating graphical and phonetic features significantly improved CNER performance.
- A higher proportion of phono-semantic characters did not necessarily yield better results.
Conclusions:
- The combination of graphical and phonetic features is effective for enhancing Chinese CNER.
- This approach offers a promising direction for processing specialized Chinese text.
- Future research can further refine feature extraction and model architectures.
