KCL-Health-NLP@CLEF eHealth 2018 Task 1: ICD-10 coding of French and Italian death certificates with character-level convolutional neural networks

TY - JOUR

T1 - KCL-Health-NLP@CLEF eHealth 2018 Task 1

T2 - 19th Working Notes of CLEF Conference and Labs of the Evaluation Forum, CLEF 2018

AU - Ive, Julia

AU - Viani, Natalia

AU - Chandran, David

AU - Bittar, André

AU - Velupillai, Sumithra

PY - 2018/1/1

Y1 - 2018/1/1

N2 - In this paper we describe the participation of the KCL-Health-NLP team in the CLEF eHealth 2018 lab, specifically Task 1: Multilingual Information Extraction-ICD10 coding. The task involves the automatic coding of causes of death in death certificates in French, Italian and Hungarian according to the ICD-10 taxonomy. Choosing to work on the two Romance languages, we treated the task as a sequence-to-sequence prediction problem. Our system has an encoder-decoder architecture, with convolutional neural networks based on character em-beddings as encoders and recurrent neural network decoders. Our hypothesis was that a character-level representation would allow our model to generalise across two genealogically related languages. Results obtained by pre-training our Italian model on the French data set confirmed this intuition. We also explored the impact of character-level features extracted from dictionary-matched ICD codes. We obtained F-measures of 0.72/0.64 and 0.78 on the French aligned/raw and Italian raw internal test data, respectively. On the blind test set released by the task organisers, our top results were 0.65/0.52 and 0.69 F-measure, respectively.

AB - In this paper we describe the participation of the KCL-Health-NLP team in the CLEF eHealth 2018 lab, specifically Task 1: Multilingual Information Extraction-ICD10 coding. The task involves the automatic coding of causes of death in death certificates in French, Italian and Hungarian according to the ICD-10 taxonomy. Choosing to work on the two Romance languages, we treated the task as a sequence-to-sequence prediction problem. Our system has an encoder-decoder architecture, with convolutional neural networks based on character em-beddings as encoders and recurrent neural network decoders. Our hypothesis was that a character-level representation would allow our model to generalise across two genealogically related languages. Results obtained by pre-training our Italian model on the French data set confirmed this intuition. We also explored the impact of character-level features extracted from dictionary-matched ICD codes. We obtained F-measures of 0.72/0.64 and 0.78 on the French aligned/raw and Italian raw internal test data, respectively. On the blind test set released by the task organisers, our top results were 0.65/0.52 and 0.69 F-measure, respectively.

KW - Convolutional neural networks

KW - Encoder-decoder architecture

KW - Recurrent neural networks

UR - http://www.scopus.com/inward/record.url?scp=85051060301&partnerID=8YFLogxK

M3 - Conference paper

AN - SCOPUS:85051060301

SN - 1613-0073

VL - 2125

JO - CEUR Workshop Proceedings

JF - CEUR Workshop Proceedings

Y2 - 10 September 2018 through 14 September 2018

ER -