Multilingual Negation Scope Resolution for Clinical Text

This repository contains code for identifying negation scopes in multilingual clinical data using a transformer-based multilingual language model, as described in the paper Multilingual Negation Scope Resolution for Clinical Text (LOUHI 2021).

The code in this repository is a re-implementation of the original code that was used for the experiments described in the paper. The original code is based on the mt-dnn framework, and it was re-implemented in order to make it less complex. The results that these models achieve are comparable to the results reported in the paper, and can be found here If you would like to obtain the code that was used for the original experiments in the paper, please contact the authors at mrkhartmann4@gmail.com

Installing required packages

Install the conda package manager following the instructions. Create a virtual env and install the required packages using

conda create --name multinegscope 
conda activate multinegscope
pip install -r requirements.txt 

Using trained models to predict negation scopes

You can use our pretrained model to predict negation scopes in your data of interest. Negation scopes are predicted in two stages: First, negation cues are identified based on a pre-defined negation cue list. Second, negation scopes are marked given the identified negation cues. In the following, we describe the steps you need to follow.

1) Identifying negation cues using a pre-defined cue list

python tag_negation_cues.py \
--cue_list ./data/cues/cues_danish.txt \
--input_file ./examples/example.jsonl

The predefined cue list should contain one cue per line (see ./data/cues/cues_danish.txt). Discontinuous cues can be indicated using the # symbol, and the code currently matches discontinuous cues with up to 6 tokens separating the different parts of the cue.

The script creates two files ./examples/example#cues.jsonl and ./examples/example#cues.html. The first can be used as input file for the negation scope resolution model, the latter visualizes the detected negation cues in red when opened in a web browser.

2) Predicting negation scopes given marked negation cues

Download the pre-trained negation scope resolution model from here and unzip it.

python predict_negscopes.py \
--model_checkpoint ./pretrained_model_v1/model.pt  \
--datapath ./examples \
--test_datasets example#cues.jsonl \
--outdir ./examples

The script creates two files ./examples/example#cues#scopes.jsonl and ./examples/example#cues#scopes.html.

Training models

Coming soon

Information about how to pre-process the data will be added here in the future. In the meantime, please contact the authors is you have any questions about dataset preprocessing.

Evaluating models

Coming soon

Contact the Authors

If you have questions, please contact the corresponding author at mrkhartmann4@gmail.com .

Our work can be cited using
@inproceedings{hartmann-sogaard-2021-multilingual,
title = "Multilingual Negation Scope Resolution for Clinical Text",
author = "Hartmann, Mareike and S{\o}gaard, Anders", booktitle = "Proceedings of the 12th International Workshop on Health Text Mining and Information Analysis",
month = apr,
year = "2021",
address = "online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2021.louhi-1.2",
pages = "7--18"}