This paper describes the system presented at the MEDDOCAN (Medical Document Anonymization) task. The system consists of a candidate generator which uses a PoS-tagger, and a candidate classi er based on a convolutional neural network which uses three channels and pretrained word embeddings to represent the sequence of words to be classi ed and its left and right context. On the MEDDOCAN Test Set, the systems achieved an F1 score of 0.9184 for subtask 1 (NER o set and entity type classi cation) and 0.9345 for subtask 2 (sensitive token detection).
This paper describes the system presented at the MEDDOCAN (Medical
Document Anonymization) task within the IberLEF 2019 initiative. MEDDOCAN is
the rst community challenge task speci cally devoted to the anonymization of
medical documents in Spanish. Although the MEDDOCAN task is structured
in two subtasks: (i ) NER o set and entity type classi cation and (ii ) sensitive
token detection, these two subtasks are approached simultaneously providing
the same output for both. A more detailed description of MEDDOCAN and the
results of systems submitted can be found in [
The anonymization or de-identi cation task can be classi ed as a named
entity recognition and classi cation problem, which has been extensively studied
from a variety of approaches ranging from rule-based systems to machine learning
algorithms. Several systems using di erent kinds of recurrent neural networks
can be found in [
The use of multichannel convolutional neural networks for sentence classi
cation was rst described in [
Jordi Porta convolutional neural network was expanded by multiple parallel convolutional neural networks with di erent kernel sizes over the same text. Here, we will use channels to represent the sequence of words to be classi ed and its left and right context. 2
The MEDDOCAN corpus is composed of 1000 clinical cases distributed over
a Training Set (500 cases), Development Set (250 cases), and a Test Set (250
cases). Each clinical case came in raw text format with its corresponding brat
annotation le [
One of these texts is shown in Fig. 1, where MEDDOCAN entities in the corresponding annotation le in Fig. 2 have been highlighted in red and section titles in blue in order to show the underlying structure of these clinical cases and the co-occurrence of entities with labels that introduce many of them (e.g., Apellidos: Garcia Prieto). Note also that, although these texts have been annotated blindly by two di erent expert annotators with 98% of inter-annotator agreement, the text contains a false positive (Testes, wrongly classi ed as FAMILIARES SUJETO ASISTENCIA) and a false negative (08022, not identi ed and classi ed as TERRITORIO). 3
The system built for MEDDOCAN consists of a candidate generator and a candidate classi er, which are depicted in Fig. 3 and described in the following paragraphs.
The candidate generation is performed in two steps: (i ) Texts are PoS-tagged with a general in-house Spanish PoS-tagger which performs sentence segmentation, tokenization, and morphosyntactic analysis. The tokenization subcomponent has been slightly modi ed to split textual elements containing hyphens and some textual elements consisting of unbroken sequences of words, numbers and punctuation without proper spacing (e.g., Edad:68 ! Edad : 68 ); and (ii ) PoS-Tagged texts are then given to an n-gram generator to provide candidates along with their textual context for all the types of entities de ned by the MEDDOCAN task. Since most of the n-grams do not correspond to entities from a linguistic point of view (they should be noun phrases), a lter is applied to reduce the number of false candidates. This lter removes n-grams with nonallowed PoS tags in n-gram initial and nal positions (punctuation, prepositions, adverbs, articles, conjunctions, etc.), n-grams with internal unpaired punctuation, and n-grams containing verbs (except past participle forms) or any of the words in a blacklist of the 15,000 most frequent non-allowed words within entities manually extracted from the MEDDOCAN texts in the Training Set.
Candidate n-grams with their context are classi ed in two steps: (i ) n-grams are initially classi ed by a three-channel convolutional neural network (TCCNN) with precomputed word embeddings (channels represent left and right context, Nombre: Marc.
Apellidos: Garcia Prieto.
CIPA: 270058.
NASS: 27 2226350 05.
Domicilio: Av. Litoral, 30, 1C .
Localidad/ Provincia: Barcelona.
CP: 08005.
NHC: 270058.
Datos asistenciales.
Fecha de nacimiento: 27-10-1968.
Pa s: Espa~na.
Edad: 47 Sexo: H.
Fecha de Ingreso: 13-12-2015.
Especialidad: androlog a.
Medico: Anna Bujons Tur NoCol: 08 08 76541.
Antecedentes: sin antecedentes patologicos de interes ni habitos toxicos. Historia Actual: Paciente varon de 47 a~nos de edad, acude a la consulta de androlog a por presentar erecciones prolongadas no dolorosas de aproximadamente 4 a~nos de evolucion tras traumatismo perineal cerrado con el manillar de una bicicleta.
Exploracion f sica:En la exploracion f sica se observan cuerpos cavernosos aumentados de consistencia, no dolorosos a la palpacion, sin palpar pulsos anomalos. Sensibilidad peneana conservada. Testes moviles en ambas bolsas escrotales y sin alteraciones.
Resumen de pruebas complementarias: Como exploraciones complementarias se le realiza ecodoppler penenano: vascularizacion cavernosa derecha aparentemente conservada; en la porcion mas proximal del cuerpo cavernoso izquierdo se observa formacion anecoica (2x1.8x1.5cm) con flujo turbulento en su interior compatible con f stula arteriovenosa (FAV) de larga duracion.
Evolucion y comentarios: Con la orientacion diagnostica de Priapismo de alto flujo se decide realizacion de Arteriograf a pudenda con anestesia local confirmandose FAV y posterior embolizacion de la misma mediante 2 coil 3x5. Tras la embolizacion el paciente e voluciona favorablemente con detumecencia peneana completa y erecciones normales. Actualmente esta asintomatico.
Diagnostico Principal: Priapismo Remitido por: Anna Bujons Tur Calle Joaquim Folguera, 3 - 5o 2a 08022 Barcelona. abujons@gmail.com
Fig. 1. MEDDOCAN Clinical Case S0004-06142006000600015-1: Raw Text
Jordi Porta and the n-gram to be classi ed); and (ii ) a nal lter is applied to discard misclassi cations and to solve overlapped classi ed candidates selecting left-most longest n-grams. This nal lter applies constraints at word and character levels on classi ed n-grams. For example, emails must contain the @ character, streets have no street type information, etc. These constraints have been created on the basis of observation of misclassi cations in the Development Set.
Despite the initial lter applied to reduce the number of candidates, 89%
of the candidates in the MEDDOCAN corpus are still non-entities, resulting in
training sets with a very unbalanced label distribution. To avoid bias to the
machine learning algorithm, labels are weighted to make them equally important
to the algorithm. Word embeddings have been trained using Word2Vec [
The topology of the TCCNN is depicted in Fig. 4. Three input channels are
de ned for processing the sequence of words to be classi ed and its left and right
context. Each channel has an input layer, limiting to 15 the length of the input
sequence, followed by an embedding layer and a one-dimensional convolutional
layer. Max pooling layers down-sample the output from the convolutional layers
and a atten layer reduces its dimension for concatenation. The concatenated
output from the three channels is processed by a dense layer and an output
layer. The code used for implementing the TCCNN is an adaptation of the code
published in [
Training and development sets were created from MEDDOCAN datasets using the candidate generator to train and evaluate the TCCNN. For each entity type, its weight was computed as n=ni where n = Pi ni and ni is the number of examples of type i. Hyperparameters of the neural network were manually explored until a peak of 0.965 accuracy was reached on the development set (corresponding to 0.992 on the training set). Channels have a capacity of 15 words, dropout rates are set to 0.005, lter windows to 32, kernel and pool sizes to 2; dense layers contain 200 units with L2 kernel regularizers set to 0.001; activation functions are ReLU except for the output layer, which is softmax. The mini-batch was 156 and the optimizer was Adam with a learning rate of 0.001. The learning curve of Fig. 5 indicates that the model ts quite well with a training-development accuracy gap of 3%. The performance of the neural net
Jordi Porta work classi er (without the nal lter) on every entity type in terms of precision and recall, ordered by F1, is shown in Table 1. There is no apparent correlation between performance and the number of examples of each entity type. The two systems submitted di er in the training corpus: System II is trained with the Training Set and System III is trained with all available data, i.e., the Training and Development Sets.
The results on Development and Test Sets in subtasks 1 and 2 are shown in Tables 2.1 and 2.2, where the e ect of the nal lter is also measured (in all cases left-most longest matches are selected in case of overlapping). Results for System III on the Development Set are not given since this set has been seen during the training. For all the experiments, the e ect of the lter is positive, increasing all gures by 0.8-3.41%. On the Test Set, Systems II and III have slightly di erent precision and recall but a similar F1 of about 0.918 for subtask 1 and 0.93 for subtask 2. 5
An initial approach to the MEDDOCAN task was to create a pattern-based system using context-free grammars to classify and generate candidates for the kind of entities de ned by the task (System I). The development of this approach results in high precision but low recall due to the di culty of modelling the span and context of some entity types. The course of the development was then changed by the n-gram generation using a blacklist of word forms combined
Spanish Medical Document Anonymization with TCCNNs
Jordi Porta with a machine learning classi er, which was previously used for named entity classi cation, resulting in a best F1 of 0.9184 for subtask 1 and 0.9345 for subtask 2. Part of the success in the tasks can be explained by the distribution of the entities within the texts, many of which are placed in elds and surrounded by labels introducing or delimiting them. A natural path for future work is to incorporate PoS tags and word character level information into new channels of the TCCNN, for ther purpose of eliminating the lter in the n-gram candidate generation and the nal lter.