A Framework for False Negative Detection in NER/NEL

Finding the false negatives of a NER/NEL system is fundamental to improve it, and is usually done by manual annotation of texts. However, in an environment with a huge volume of unannotated texts (e.g. a hospital) and a low frequency of positives (e.g. a mention of a particular disease in the clinical notes) the task becomes very inefficient. This paper presents a framework to tackle this problem: given an existing NER/NEL system, we propose a technique consisting of using text similarity search to rank texts by probability of containing false negatives of a given concept, using as a query those texts where the existing NER/NEL system has found positives of this concept. We formulate text similarity as a function of shared medical entities between texts, and we re-purpose an existing public dataset (CodiEsp) to propose an evaluation strategy.