Clinical Text Prediction with Numerically Grounded Conditional Language\n Models

Assisted text input techniques can save time and effort and improve text\nquality. In this paper, we investigate how grounded and conditional extensions\nto standard neural language models can bring improvements in the tasks of word\nprediction and completion. These extensions incorporate a structured knowledge\nbase and numerical values from the text into the context used to predict the\nnext word. Our automated evaluation on a clinical dataset shows extended models\nsignificantly outperform standard models. Our best system uses both\nconditioning and grounding, because of their orthogonal benefits. For word\nprediction with a list of 5 suggestions, it improves recall from 25.03% to\n71.28% and for word completion it improves keystroke savings from 34.35% to\n44.81%, where theoretical bound for this dataset is 58.78%. We also perform a\nqualitative investigation of how models with lower perplexity occasionally fare\nbetter at the tasks. We found that at test time numbers have more influence on\nthe document level than on individual word probabilities.\n