--- license: cc-by-4.0 --- # The Tokenizer for Clinical Cases Written in Spanish ## Introduction This repository contains the tokenization model trained using the SPACCC_TOKEN corpus (https://github.com/PlanTL-SANIDAD/SPACCC_TOKEN). The model was trained using the 90% of the corpus (900 clinical cases) and tested against the 10% (100 clinical cases). This model is a great resource to tokenize biomedical documents, specially clinical cases written in Spanish. This model was created using the Apache OpenNLP machine learning toolkit (https://opennlp.apache.org/), with the release number 1.8.4, released in December 2017. This repository contains the training set, testing set, Gold Standard. ## Prerequisites This software has been compiled with Java SE 1.8 and it should work with recent versions. You can download Java from the following website: https://www.java.com/en/download The executable file already includes the Apache OpenNLP dependencies inside, so the download of this toolkit is not necessary. However, you may download the latest version from this website: https://opennlp.apache.org/download.html The library file we have used to compile is "opennlp-tools-1.8.4.jar". The source code should be able to compile with the latest version of OpenNLP, "opennlp-tools-*RELEASE_NUMBER*.jar". In case there are compilation or execution errors, please let us know and we will make all the necessary updates. ## Directory structure
exec/
  An executable file that can be used to apply the tokenization to your documents.
  You can find the notes about its execution below in section "Usage".

gold_standard/
  The clinical cases used as gold standard to evaluate the model's performance.

model/
  The tokenizationint model, "es-tokenization-model-spaccc.bin", a binary file.

src/
  The source code to create the model (CreateModelTok.java) and evaluate it (EvaluateModelTok.java).
  The directory includes an example about how to use the model inside your code (Tokenization.java).
  File "abbreviations.dat" contains a list of abbreviations, essential to build the model.

test_set/
  The clinical cases used as test set to evaluate the model's performance.

train_set/
  The clinical cases used to build the model. We use a single file with all documents present in
  directory "train_set_docs" concatented.

train_set_docs/
  The clinical cases used to build the model. For each record the sentences are already splitted.

## Usage The executable file *Tokenizer.jar* is the program you need to tokenize the text in your document. For this program, two arguments are needed: (1) the text file to tokenize, and (2) the model file (*es-tokenization-model-spaccc.bin*). The program will display all tokens in the terminal, with one token per line. From the `exec` folder, type the following command in your terminal:
$ java -jar Tokenizer.jar INPUT_FILE MODEL_FILE
## Examples Assuming you have the executable file, the input file and the model file in the same directory:
$ java -jar Tokenizer.jar file.txt es-tokenizer-model-spaccc.bin
## Model creation To create this tokenization model, we used the following training parameters (class *TrainingParameters* in OpenNLP) to get the best performance: - Number of iterations: 1500. - Cutoff parameter: 4. - Trainer type parameter: *EventTrainer.EVENT_VALUE*. - Algorithm: Maximum Entropy (*ModelType.MAXENT.name()*). Meanwhile, we used the following parameters for the tokenizer builder (class *TokenizerFactory* in OpenNLP) to get the best performance: - Language code: *es* (for Spanish). - Abbreviation dictionary: file "abbreviations.dat" (included in the `src/` directory). - Use alphanumeric optimization: false - Alphanumeric pattern: null ## Model evaluation After tuning the model using different values for each parameter mentioned above, we got the best performance with the values mentioned above. | | Value | | ----------------------------------------: | :------ | | Number of tokens in the gold standard | 38247 | | Number of tokens generated | 38227 | | Number of words correctly tokenized | 38182 | | Number of words wrongly tokenized | 35 | | Number of tokens missed | 30 | | **Precision** | **99.88%** | | **Recall** | **99.83%** | | **F-Measure** | **99.85%**| Table 1: Evaluation statistics for the tokenization model. ## Contact Ander Intxaurrondo (ander.intxaurrondo@bsc.es) ## License Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License. Copyright (c) 2018 Secretaría de Estado para el Avance Digital (SEAD)