Building a Multimodal Entity Linking Dataset From Tweets

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 413–421 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 413 Alignment Annotation for Clinic Visit Dialogue to Clinical Note Sentence Language Generation Wen-wai Yim†, Meliha Yetisgen*, Jenny Huang‡, Micah Grossman† † Augmedix, Inc., *University of Washington, ‡ University of California, Los Angeles † 1161 Mission St 210, San Francisco 94103, ∗850 Republican St, Seattle WA 98109, ‡ 405 Hilgard Avenue, Los Angeles, CA 90095 wenwai.yim@augmedix.com, melihay@uw.edu, jyh08@ucla.edu, micah.grossman@augmedix.com Abstract For every patient’s visit to a clinician, a clinical note is generated documenting their medical conversation, including complaints discussed, treatments, and medical plans. Despite advances in natural language processing, automating clinical note generation from a clinic visit conversation is a largely unexplored area of research. Due to the idiosyncrasies of the task, traditional methods of corpus creation are not effective enough approaches for this problem. In this paper, we present an annotation methodology that is content- and technique- agnostic while associating note sentences to sets of dialogue sentences. The sets can further be grouped with higher order tags to mark sets with related information. This direct linkage from input to output decouples the annotation from specific language understanding or generation strategies. Here we provide data statistics and qualitative analysis describing the unique annotation challenges. Given enough annotated data, such a resource would support multiple modeling methods including information extraction with template language generation, information retrieval type language generation, or sequence to sequence modeling. Keywords: Corpus, Dialogue 1. Introduction Two trends drive significant interest in automating the pro- cess of medical documentation: a growing shortage of clin- icians in the United States and a rise in clinician burnout rates due to health information technology-related stress (Gidwani et al., 2017; AAMC, 2019; Gardner et al., 2019). Clinicians today are responsible for more than just the well- being of their patients. They must also complete docu- mentation on their time present with the patient, diagnoses made, and treatments prescribed in the electronic medi- cal record. Medical scribes, who can assist clinicians with completing medical documentation, are one way clinicians can unburden themselves from documentation responsibil- ities. But the cost of employing a medical scribe, estimated between $49K (onsite) and $23K (virtual) a year (Brady and Shariff, 2013), is prohibitive. Despite recent natural language processing (NLP) advance- ments, such as improved speech to text, deep learning NLP modeling, and greater availability of clinical NLP re- sources, the task of converting a clinic visit conversation into its corresponding clinical note remains challenging. True comprehension of the clinical situation discussed in a visit requires many difficult aspects of language under- standing and generation, such as summarizing over mul- tiple statements and question-answer responses. Clinical note generation also depends on input from outside fac- tors, e.g. electronic medical record data, templates, and pa- tients’ reported visit complaints. When a scribe is present, intake information can come from both a clinic visit con- versation dialogue as well as direct communication with …

Content truncated to 500 words