Most research on biomedical named entity recognition has focused on English texts, e.g., MEDLINE abstracts. However, recent years have also seen significant growth of biomedical publications in other languages. For example, the Chinese Biomedical Bibliographic Database has collected over 3 million articles published after 1978 from 1600 Chinese biomedical journals. We present here a Conditional Random Field (CRF) based system for recognizing biomedical named entities in Chinese texts. Viewing Chinese sentences as sequences of characters, we trained and tested the CRF model using a manually annotated corpus containing 106 research abstracts (481 sentences in total). The features we used for the CRF model include word segmentation tags provided by a segmenter trained on newswire corpora, and lists of frequent characters gathered from training data and external resources. Randomly selecting 400 sentences for training and the rest for testing, our system obtained an 68.60% F-score on average, significantly outperforming the baseline system (F-score 60.54% using a simple dictionary match). This suggests that statistical approaches such as CRFs based on annotated corpora hold promise for the biomedical NER task in Chinese texts.
展开▼
