earticle

영어

Lexical and syntactic analysis, including word segmentation, part-of-speech (POS) tagging, shallow parsing and full parsing, are essential for medical language processing (MLP). However, research on full parsing, even shallow parsing and POS tagging for Chinese electronic medical record (CEMR), has not been carried out because of the lack of annotated corpus on CEMR. In this paper, we built a corpus of 5,024 sentences from CEMR with word segmentation, POS tags and phrase tags, of them, 2,553 are annotated as full parsing trees. Inter-annotator agreement results: Chinese word segmentation (97.56%), POS tagging (93.34%), shallow parsing (96.5%), full parsing (91.22%). A lexical and syntactic analysis system for CEMR is developed and evaluated based on above corpus. Of its components, we proposed a joint model for word segmentation and POS tagging with the transformation-based error-driven model as correction postprocessing to alleviate the problem of error accumulation, the F1-score of word segmentation and POS tagging were 94.39% and 93.2%, respectively. A shallow parsing model under the framework of group learning we proposed was developed, which enriched word features by word embedding from large unlabeled CEMRs and achieved the F1-score of 96.3%. At last, we presented a state-of-art full parser combining the Berkeley parser and the Stanford parser to outperform the best single parser by 3.68%. The evaluation results show a substantial benefit to statistical machine learning models from the annotated CEMR. These works are the foundation for natural language processing (NLP) technologies applied to CEMR.