Part-of-speech tagging for clinical text: wall or bridge between institutions?

Part-of-speech (POS) tagging is a fundamental step required by various NLP systems. The training of a POS tagger relies on sufficient quality annotations. However, the annotation process is both knowledge-intensive and time-consuming in the clinical domain. A promising solution appears to be for institutions to share their annotation efforts, and yet there is little research on associated issues. We performed experiments to understand how POS tagging performance would be affected by using a pre-trained tagger versus raw training data across different institutions. We manually annotated a set of clinical notes at Kaiser Permanente Southern California (KPSC) and a set from the University of Pittsburg Medical Center (UPMC), and trained/tested POS taggers with intra- and inter-institution settings. The cTAKES POS tagger was also included in the comparison to represent a tagger partially trained from the notes of a third institution, Mayo Clinic at Rochester. Intra-institution 5-fold cross-validation estimated an accuracy of 0.953 and 0.945 on the KPSC and UPMC notes respectively. Trained purely on KPSC notes, the accuracy was 0.897 when tested on UPMC notes. Trained purely on UPMC notes, the accuracy was 0.904 when tested on KPSC notes. Applying the cTAKES tagger pre-trained with Mayo Clinic's notes, the accuracy was 0.881 on KPSC notes and 0.883 on UPMC notes. After adding UPMC annotations to KPSC training data, the average accuracy on tested KPSC notes increased to 0.965. After adding KPSC annotations to UPMC training data, the average accuracy on tested UPMC notes increased to 0.953. The results indicated: first, the performance of pre-trained POS taggers dropped about 5% when applied directly across the institutions; second, mixing annotations from another institution following the same guideline increased tagging accuracy for about 1%. Our findings suggest that institutions can benefit more from sharing raw annotations but less from sharing pre-trained models for the POS tagging task. We believe the study could also provide general insights on cross-institution data sharing for other types of NLP tasks.