The code in this repository may be used to reproduce the author ID results found in the article:

A Deep Metric Learning Approach to Account Linking. NAACL (2021)
Aleem Khan, Elizabeth Fleming, Noah Schofield, Marcus Bishop, and Nicholas Andrews
[arxiv] [aclweb] [bib]

We provide instructions to use the preprocessed data below. To access the raw dataset (~23GB), please complete this form: https://forms.gle/TRga2LANBf4Zfo6W9

Please cite the paper above if you find this code or ideas from the paper useful in your own research.

For the latest related work, please see: https://cs.jhu.edu/~noa

In order to run the software in this repository, the packages listed in requirements.txt must be installed. One way to do this is to make a conda environment (conda create --prefix ./env python=3.8), activate it (conda activate ./env), and run pip install -r requirements.txt.

Then, include the location of this repository in your Python path:

export PYTHONPATH=$PYTHONPATH:${PATH_TO_THIS_REPOSITORY}

will enable you to import the aid (author ID) package.

To run an experiment, first fetch and unpack the data files

https://storage.googleapis.com/naacl21_account_linking/1mil.tar.gz
https://storage.googleapis.com/naacl21_account_linking/dev_queries.tar.gz
https://storage.googleapis.com/naacl21_account_linking/dev_targets.tar.gz
https://storage.googleapis.com/naacl21_account_linking/test_queries.tar.gz
https://storage.googleapis.com/naacl21_account_linking/test_targets.tar.gz

and save them onto your filesystem, specifically into separate subdirectories train, dev, and test, to prevent name conflicts. Unpack the files and then run the json2tf.py script to produce sharded protocol buffer files for training. For example:

python json2tf.py --json /path/to/unpacked/json --tf /path/to/output/tfrecords --config /path/to/reddit/json/config`

NOTE: This will take several hours for the 1mil.tar.gz files.

Next, update the JOBS_DIR variable in run_training.sh to point to the location on your filesystem where output files should be written.

Finally, to run the experiment corresponding with one of the configuration files, say, the model availing of all features, using triplet loss, and varying episode lengths between one and sixteen posts, run the command:

./run_training.sh full_model.cfg --train_records=<TRAINING DATA> --train_tfrecord_path=<VALIDATION QUERIES> --valid_tfrecord_Path=<VALIDATION TARGETS>

NOTE: When specifying the data location, use a wildcard to specify the set of all sharded files. For example, dev/queries*, which you might need to quote to prevent your shell from expanding it.

Here, run_training.sh is a wrapper for the main trainer scripts/fit.py and its argument is a configuration file, which consists of a sequence of command-line arguments, one per line. These will override any defaults specified in scripts/fit.py. You may also provide additional command-line arguments at the end of the command above. If you run the run_training.sh script on a machine with a GPU, the script should automatically recognize and use a CUDA device.

The pre-trained checkpoints below may be used to reproduce experimental results from the paper:

Checkpoints:

https://storage.googleapis.com/naacl21_account_linking/full_checkpoint.tar.gz
https://storage.googleapis.com/naacl21_account_linking/text_time_checkpoint.tar.gz

Pass the rank or link flags to scripts/fit.py to evaluate the performance of these checkpoints. See expts/reddit/test.sh.