Figure 2.

User interface: (A) The grid view shows the extracted variables in columns and individual documents in rows, providing an overview of NLP results. Below the grid, we have statistics on the active variable, with (B) the distribution of classifications for the selected variable and (C) the list of top indicators for that variable aggregated across all documents in the dataset. (D) Indicators from the active report are shown on the right. (E) The document view shows the full text of the patient reports, with the indicator terms highlighted. (F) Feedback can be sent using the control bar on the top or a right-click context menu. (G) The Word Tree view provides the ability to search and explore word sequence patterns found across the documents in the corpus, and to provide feedback that will be used to retrain NLP models. In this figure, we built the tree by searching for the word “biopsy” and then drilled down upon the node “hot.” The word tree now contains all the sentences in the dataset with the phrase “hot biopsy.” This allows the user to get an idea of all the scenarios in which “hot biopsy” has been used. Hovering over different nodes in the tree will highlight specific paths in the tree with the selected term. (H) The retrain view lists user-provided feedback, including any potential inconsistencies, and specifies changes in variable assignments due to retraining. In the example above, the user has selected a text span documenting “informed consent” in a report. However, the user also labeled the report incorrectly, possibly in error. NLPReViz points this out as conflicting feedback.