CDV – Contextual Discourse Vectors

Code and data for:

Learning Contextualized Document Representations for Healthcare Answer Retrieval. Sebastian Arnold, Betty van Aken, Paul Grundmann, Felix A. Gers and Alexander Löser. The Web Conference 2020 (WWW'20). ACM, 2020: 1332–1343.

This code is based on the TeXoo Framework and Eclipse Deeplearning4j.

Preparing datasets

We provide full training and evaluation data from Wikipedia (WikiSectionQA) and annotations that extend MedQuAD and HealthQA datasets. You have to run convert.sh inside the datasets' directories in order to prepare the json files from the original sources. Please see the instructions in the data directory for more details.

Training a model

• run bin/train-cdv

TeXoo: train contextualized discourse vectors (CDV)
 -a,--aspect <arg>        path to the aspect embedding
 -b,--balancing           use class balancing during training
 -d,--datasetname <arg>   name of the data set, e.g. wd_disease
 -e,--entity <arg>        path to the entity embedding
 -i,--dataset <arg>       path to the WikiSection training dataset
 -m,--modelname <arg>     model name
 -o,--output <arg>        path to create the output folder in
 -s,--search <arg>        search path for pre-trained word embeddings
 -u,--ui                  enable training UI
 -w,--wordemb <arg>       path to a pretrained word embedding

• Examples:

bin/train-cdv -m "mymodel" -i data/Train/wd_disease_train.json -d wd_disease -s models/common -w models/common/en_disease_skipgram.bin -e models/ENC-E@wd_disease+ft-lstm+128 -a models/ENC-A@wd_disease+ft-lstm+128 -o models

Hyperparameters are configured in the source file. See below how to get access to pre-trained ENC-E / ENC-A embeddings.

Running the evaluation

• run bin/evaluate-cdv

TeXoo: evaluate CDV answer retrieval
 -a,--aspect <arg>    optional path to a CDV single-task aspect model
 -d,--dataset <arg>   path to the evaluation dataset (json)
 -e,--entity <arg>    optional path to a CDV single-task entity model
 -m,--model <arg>     path to the pre-trained CDV multi-task model
 -p,--path <arg>      search path to sentence embedding models (if not
                      provided by the model itself)

• Example:

bin/evaluate-cdv -m models/CDV@wd_disease+avg-fasttext -p models/common -d data/WikiSectionQA/WikiSectionQA_test.json

Please contact sarnold(at)beuth-hochschule(dot)de to get access to the pre-trained CDV+avg-fasttext model.

Cite

@inproceedings{arnold2020learning,
  author = {Arnold, Sebastian and {van Aken}, Betty and Grundmann, Paul and Gers, Felix A. and L{\"o}ser, Alexander},
  title = {Learning {{Contextualized Document Representations}} for {{Healthcare Answer Retrieval}}},
  booktitle = {Proceedings of The Web Conference 2020 (WWW '20)},
  year = {2020},
  isbn = {9781450370233},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3366423.3380208},
  doi = {10.1145/3366423.3380208},
  booktitle = {Proceedings of The Web Conference 2020},
  pages = {1332–1343},
  location = {Taipei, Taiwan},
  series = {WWW ’20}
}