Text Segmentation of Digitized Clinical Texts

Text segmentation of digitized clinical texts Cyril Grouin LIMSI, CNRS, Universit´e Paris-Saclay, F-91405 Orsay cyril.grouin@limsi.fr Abstract In this paper, we present the experiments we made to recover the original page layout structure into two columns from layout damaged digitized files. We designed several CRF-based approaches, either to identify column separator or to classify each token from each line into left or right columns. We achieved our best results with a model trained on homogeneous corpora (only files composed of 2 columns) when classifying each token into left or right columns (overall F-measure of 0.968). Our experiments show it is possible to recover the original layout in columns of digitized documents with results of quality. Keywords: Digitized texts; Document layout; Natural Language Processing 1. Introduction 1.1. Digitization and page decomposition The digitization process generally involves several steps (Randriamasy and Vincent, 1994): (i) page decom- position in order to recognize structural and logical units (e.g., blocks, columns, lines) within a page, (ii) characters recognition using either Optical Character Recognition (OCR) or Handwritten Text Recognition (HTR), and (iii) format and layout analysis so as to produce the final layout of a digitized document. The performances of the page decomposition step is crucial for the final output. The delimitation of text blocks, columns, and lines is achieved using the spatial properties of the elements on a page (Ha et al., 1995). Sylwester and Seth (1995) designed a specific algorithm for the bloc segmentation task, based on the XY Cut1 algorithms series, which relies on costs defined for each possible block decomposition from the image. More recently, new techniques have been designed to deal with complex document layout (multi-columns, noisy ar- eas, ornamental characters, etc.) or for a particular domain. Specific pre- and post-processing techniques tailored for a given domain tend to improve text extraction from digitized text, as done by Xu et al. (2008) for the biomedical do- main. Nikolaou et al. (2010) proposed an algorithm for an- cient machine-printed documents (e.g., old books from the 18th century) based on the run length smoothing algorithm (RLSA). They improved the results achieved by the RLSA algorithm for both line, word and character recognitions. Kaur et al. (2013) demonstrated that a pre-processing stage can improve the page decomposition task. Alternatively, post-processing techniques are also used to correct OCR errors which constitute obstacles to many further NLP pro- cesses. Those post-processing techniques aim at identify- ing tables within the text (Kieninger, 1998; Ng et al., 1999) or correcting tokenization errors (Furrer, 2013). Since performances of page decomposition also depends on the metrics used, Shafait et al. (2007) compared six metrics to evaluate the page decomposition functions from OCR systems in order to select the metric that provides the best decomposition outputs. 1XY Cut or top-down algorithms represent the whole docu- ment as the root of a tree. Each final segmentation is represented as a leaf in this tree. 1.2. Recovering the original document layout Nevertheless, the global layout of digitized documents can be lost, especially frontiers of …

Content truncated to 500 words