Background: Competency-based medical education relies heavily on high-quality narrative reflections and feedback within workplace-based assessments. However, evaluating these narratives at scale remains a significant challenge. Objective: This study aims to develop and apply natural language processing (NLP) models to evaluate the quality of resident reflections and faculty feedback documented in Entrustable Professional Activities (EPAs) on Taiwan’s nationwide Emyway platform for otolaryngology residency training. Methods: This 4-year cross-sectional study analyzes 300 randomly sampled EPA assessments from 2021 to 2025, covering a pilot year and 3 full implementation years. Two medical education experts independently rated the narratives based on relevance, specificity, and the presence of reflective or improvement-focused language. Narratives were categorized into 4 quality levels—effective, moderate, ineffective, or irrelevant—and then dichotomized into high quality and low quality. We compared the performance of logistic regression, support vector machine, and bidirectional encoder representations from transformers (BERT) models in classifying narrative quality. The best performing model was then applied to track quality trends over time. Results: The BERT model, a multilingual pretrained language model, outperformed other approaches, achieving 85% and 92% accuracy in binary classification for resident reflections and faculty feedback, respectively. The accuracy for the 4-level classification was 67% for both. Longitudinal analysis revealed significant increases in high-quality reflections (from 70.3% to 99.5%) and feedback (from 50.6% to 88.9%) over the study period. Conclusions: BERT-based NLP demonstrated moderate-to-high accuracy in evaluating the narrative quality in EPA assessments, especially in the binary classification. While not a replacement for expert review, NLP models offer a valuable tool for monitoring narrative trends and enhancing formative feedback in competency-based medical education.

In this paper, we present a observation that a word with some special sense should appear in specific category of articles, and this appearance will “take” the word to its Wiki sense [1] by Wikipedia link graph. This property can help to solve word sense disambiguation problem using Wikipedia.

This study utilizes the danmaku on the Bilibili platform as the research subject to examine how their characteristics vary according to the nature or focus of English teaching videos. By employing social network analysis, the study reveals distinctive features in danmaku. For videos categorized under linguistic knowledge (phonetics, vocabulary, and grammar), the danmaku comments predominantly center around topics such as phonetics, vocabulary, and grammar. Conversely, in videos categorized under language skills (listening, speaking, reading and writing), the danmaku comments primarily reflect a vocabulary review for three of the four skills, with only the listening skill showing slight deviations. This underscores the centrality of vocabulary in skill-oriented videos. The findings highlight the unique role of danmaku in distinguishing between knowledge and skills within the context of English teaching videos.

This paper describes the process of extracting and filtering Wikipedia revision history as a resource for grammatical error correction (GEC). Edits in Wikipedia revision history vary widely, including grammatical error corrections, information supplements, format amendments, and even vandalism. To extract only GEC-related revisions, we use an automated error annotation toolkit, ERRANT ¹ , and extend it to process large data in parallel efficiently. With error-type analysis, we can then identify GEC-related edits and omit other unrelated edits (i.e., only the correction parts are reserved). The resulting corpus is — to our knowledge — the largest publicly available corpus of parallel possibly erroneous and correct sentences with error type labels.