The NCBO Annotator 'automatically processes a piece of raw text to annotate (or tag) it with relevant ontology concepts and return the annotations' (Jonquet, 2009).
For MDDict, words were harvested from three sources: a free dictionary of contemporary German [ 14], a word list created out of raw text extracted from an old CD-ROM version of a medical dictionary [ 15], and medical texts and forum postings from a patient-centered website [ 16].
The proposed method includes NLP tasks for text preprocessing and learning word representation features from a large amount of raw text data, in addition to the word, the word n-gram, the character n-gram, and the traditional orthographic information (baseline features) for feature extraction, and applies CRF for token labeling.
Instead, they can be constructed for any language or domain, as long as a reasonable amount of raw text in electronic format is available.
As Fig. 1 shows, step 5 may give output to a secondary use of the raw text and data file to produce a more comprehensive and publishable output.
In our implementation, for a given split sentence, if its previous line in the original raw text was an empty line, the value of the layout feature was 1, or otherwise it was 0. Take the line starting with the section "Name:" of Table 1 as an example.
Other actions that have helped keep the amount of data low on the fast, but expensive parallel storage system at UPPMAX have been the implementation of more strict policies for allowances, cleaning up of temporary data, compressing files in inefficient file formats like raw text, and an increased use of the SweStore national storage.
This means that we kept the annotations separate from the actual text (the information on the location of the entity mentions is stored in a different file from the actual raw text).
To obtain reliable results that are less sensitive to noise, we run a few preprocessing steps on the raw text accompanying crime events including removal of so-called stop-words (see e.g., Rajman and Besançon 1998).
