Zafar Atif, Mamlin Burke, Perkins Susan, Belsito Anne M, Overhage J Marc, McDonald Clement J
School of Medicine, Regenstrief Institute, Indiana University, 1001 West 10th Street, RG5 Indianapolis, IN 46202, USA.
Int J Med Inform. 2004 Sep;73(9-10):719-30. doi: 10.1016/j.ijmedinf.2004.05.008.
To (1) discover the types of errors most commonly found in clinical notes that are generated either using automatic speech recognition (ASR) or via human transcription and (2) to develop efficient rules for classifying these errors based on the categories found in (1). The purpose of classifying errors into categories is to understand the underlying processes that generate these errors, so that measures can be taken to improve these processes.
We integrated the Dragon NaturallySpeaking v4.0 speech recognition engine into the Regenstrief Medical Record System. We captured the text output of the speech engine prior to error correction by the speaker. We also acquired a set of human transcribed but uncorrected notes for comparison. We then attempted to error correct these notes based on looking at the context alone. Initially, three domain experts independently examined 104 ASR notes (containing 29,144 words) generated by a single speaker and 44 human transcribed notes (containing 14,199 words) generated by multiple speakers for errors. Collaborative group sessions were subsequently held where error categorizes were determined and rules developed and incrementally refined for systematically examining the notes and classifying errors.
We found that the errors could be classified into nine categories: (1) announciation errors occurring due to speaker mispronounciation, (2) dictionary errors resulting from missing terms, (3) suffix errors caused by misrecognition of appropriate tenses of a word, (4) added words, (5) deleted words, (6) homonym errors resulting from substitution of a phonetically identical word, (7) spelling errors, (8) nonsense errors, words/phrases whose meaning could not be appreciated by examining just the context, and (9) critical errors, words/phrases where a reader of a note could potentially misunderstand the concept that was related by the speaker.
A simple method is presented for examining errors in transcribed documents and classifying these errors into meaningful and useful categories. Such a classification can potentially help pinpoint sources of such errors so that measures (such as better training of the speaker and improved dictionary and language modeling) can be taken to optimize the error rates.
（1）找出在使用自动语音识别（ASR）或人工转录生成的临床记录中最常见的错误类型；（2）根据（1）中发现的类别制定有效的错误分类规则。将错误分类的目的是了解产生这些错误的潜在过程，以便采取措施改进这些过程。
我们将Dragon NaturallySpeaking v4.0语音识别引擎集成到Regenstrief医疗记录系统中。在说话者进行纠错之前，我们捕获了语音引擎的文本输出。我们还获取了一组人工转录但未校正的记录用于比较。然后，我们仅根据上下文尝试对这些记录进行纠错。最初，三位领域专家独立检查了由一位说话者生成的104份ASR记录（包含29,144个单词）和由多位说话者生成的44份人工转录记录（包含14,199个单词）中的错误。随后举行了协作小组会议，确定错误类别，制定并逐步完善规则，以便系统地检查记录并对错误进行分类。
我们发现错误可分为九类：（1）由于说话者发音错误导致的发音错误；（2）因缺少术语而产生的词典错误；（3）由单词适当时态误识别导致的后缀错误；（4）添加的单词；（5）删除的单词；（6）由同音词替换导致的同音异形词错误；（7）拼写错误；（8）无意义错误，即仅通过检查上下文无法理解其含义的单词/短语；（9）关键错误，即记录的读者可能会误解说话者所传达概念的单词/短语。
提出了一种简单的方法来检查转录文档中的错误，并将这些错误分类为有意义且有用的类别。这种分类可能有助于确定此类错误的来源，以便采取措施（如对说话者进行更好的培训以及改进词典和语言建模）来优化错误率。
