Effect of Preprocessing for Distributed Representations: Case Study of Japanese Radiology Reports

A radiology report is a medical document based on an examination image in a hospital. However, the preparation of this report is a burden on busy physicians. To support them, a retrieval system of past documents to prepare radiology reports is required. In recent years, distributed representation has been used in various NLP tasks and its usefulness has been demonstrated. However, there is not much research about Japanese medical documents that use distributed representations. In this study, we investigate preprocessing on a retrieval system with a distributed representation of the radiology report, as a first step. As a result, we confirmed that in word segmentation using Morphological analyzer and dictionaries, medical terms in radiology reports are not handled as long nouns, but are more effective as shorter nouns like subwords. We also confirmed that text segmentation by SentencePiece to obtain sentence distributed representation reflects more sentence characteristics. Furthermore, by removing some phrases from the radiology report based on frequency, we were able to reflect the characteristics of the document and avoid unnecessary high similarity between documents. It was confirmed that preprocessing was effective in this task.