Obtaining Prebuilt Dictionaries
The dictionaries and models used during annotation indeed are the cornerstone of quality for your results. The install instructions show you how to get the separately-downloadable ctakes-resources archive (which is not itself released by the Apache Software Foundation) that you need to run most of cTAKES. Those resources include:
- An RxNorm_index database (a Lucene index): Contains drug names from RxNorm.
- The OrangeBook: If you are not using the drug NER pipeline, the Orange Book is used to filter out what it found in RxNorm so that only things in both RxNorm and Orange Book are annotated. If you use Drug NER, Orange Book filtering is bypassed.
- A UMLS database (using two hsqldb tables): Contains terms for anatomical sites, procedures, signs/symptoms, and disorders/diseases from SNOMED-CT, NCI Thesaurus, MeSH, and ICD-9 (umls_ms_2011ab) which have been tokenized by cTAKES.
- The full LVG: From the lexical tools provided by the NLM for word normalization. Used to match similar words, for example the plural and singular forms of a word.
Building Your Own Dictionaries
The UMLS dictionaries within the ctakes-resources archive might not match your underlying data completely. You might require other local terms, etc. To install customized dictionaries for RxNorm, SNOMED-CT, or other vocabularies that are available through the UMLS, see the following posts on the cTAKES forums:
As of Apache cTAKES 3.1, the models needed to run cTAKES are included with the convenience binaries.
Building your own Models
You may not need to use any models other than those provided with Apache cTAKES, however they have been trained on a specific set of text (a corpus) which might not match the characteristics of your text. If you want to build or train your own models, please read the cTAKES 3.1 Component Use Guide, particularly: