PAN 2020/2021 Authorship Verification Source Code

AdHominem describes a Siamese network topology for (binary) authorship verification, also known as pairwise (1:1) forensic text comparison. This implementation was developed for the PAN 2020 and PAN 2021 authorship verification shared tasks. It represents a hierarchical fusion of three well-known approaches into a single end-to-end learning procedure: A deep metric learning framework at the bottom aims to learn a pseudo-metric that maps a document of variable length onto a fixed-sized linguistic embedding vetor (LEV). Next, we incorporate a probabilistic layer to perform Bayes factor scoring (BFS) in the learned metric space. Finally, an uncertainty adaptation layer (UAL) is used to find and correct possible wrongly classified trials of the BFS layer, to model its noise behavior, and to return re-calibrated posteriors. Our framework additionally includes an out-of-distribution detector (O2D2) for defining non-responses. As with most deep-learning approaches, the success of the proposed architecture depends heavily on the availability of a large collection of text samples with many examples of representative variations in writing style. The size of the train set has been increased synthetically by dissembling all predefined document pairs and re-sampling new same-author and different-author pairs in each epoch. To evaluate our mode, we used the offical evaluation script taken from the official PAN repository and the calibration metrics provided here.

The random disjoint dataset splits are summarized in the figure below and vary. Altogether, the following datasets have been involved in the PAN 2021 shared task, to train the model components, tune the hyper-parameter and for testing:

• The training set was employed for the first stage, i.e., to train the DML, BFS and UAL components simultaneously. During training we re-sampled the pairs epoch-wise such that all documents contribute equally to the neural network training in each epoch.
• The calibration set has been used for the second stage, i.e., to train (calibrate) the O2D2 model. During training, we again re-sampled the pairs in each epoch.
• The purpose of the validation set is to tune the hyper-parameters of the O2D2 stage and to report the final evaluation metrics.
• The development set was used to tune the hyper-parameters during the training of the first stage. This dataset contains documents from the calibration and validation sets. However, due to the pair re-sampling strategy, it does not represent a union of the calibration and validation sets.
• The official test set, which was not publicly available, has been used to test our submission and to compare it with the proposed frameworks of all other participants.
• The validation and development sets only contain same-author/different-fandoms and different-authors/same-fandom pairs, for reasons discussed in the paper. The pairs of these sets are sampled once and then kept fixed.

A single dataset split and pre-trained models can be found here.

In the first stage, we obtained the following results (development set, epoch 32):

• PAN 20/21 metrics:

  	AUC	c@1	f_05_u	F1	Brier	overall
  DML	0.971	0.913	0.926	0.911	0.933	0.931
  BFS	0.971	0.912	0.929	0.909	0.933	0.931
  UAL	0.971	0.913	0.929	0.910	0.934	0.931

• Calibration metrics:

  	avg_acc	avg_conf	ECE	MCE
  DML	0.913	0.894	0.0192	0.063
  BFS	0.912	0.907	0.0065	0.043
  UAL	0.913	0.914	0.0045	0.019

In the second stage, we obtained the following results (validation set, epoch 40, non-responses: 7.9%):

• PAN 20/21 metrics:

  	AUC	c@1	f_05_u	F1	Brier	overall
  DML	0.963	0.906	0.917	0.903	0.926	0.923
  BFS	0.964	0.906	0.922	0.903	0.927	0.924
  UAL	0.965	0.906	0.922	0.903	0.928	0.925
  O2D2	0.965	0.927	0.913	0.929	0.927	0.932

• Calibration metrics:

  	avg_acc	avg_conf	ECE	MCE
  DML	0.906	0.889	0.0256	0.066
  BFS	0.906	0.902	0.0160	0.038
  UAL	0.906	0.909	0.0089	0.044
  O2D2	0.887	0.900	0.0217	0.154

After inference, you can plot the reliability diagrams as described in this paper

Please, feel free to send any comments or suggestions! (benedikt.boenninghoff[at]rub.de)

Installation

1. Clone git repository:

   $ git clone https://github.com/boenninghoff/pan_2020_2021_authorship_verification.git
$ cd pan_2020_2021_authorship_verification/

2. Create Python 3.6 environment using e.g. Anaconda:

   $ conda create -n PAN2021 python=3.6
$ conda activate PAN2021

3. Install libraries (Tensorflow, Spacy, etc.):

   $ pip install -r requirements.txt

4. Create folder, download pre-trained word embeddings (Fasttext) and the Spacy tokenizer:

   $ mkdir data_original && cd data_original
$ wget https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.en.300.bin.gz
$ gunzip cc.en.300.bin.gz
$ python -m spacy download en_core_web_lg

5. Download the PAN AV 2020 Dataset. We assume that the pairs/truth files for the large dataset are provided as follows:

   • /data_original/pairs.jsonl
   • /data_original/labels.jsonl

Text preprocessing, model training and inference

1. Run the preprocessing scripts. The dataset splits are described here. All preprocessing results are stored in the folder data_preprocessed/

   $ cd preprocessing
$ python step1_parse_and_split.py
$ python step2_preprocess.py
$ python step3_count.py
$ python step4_make_vocabularies.py
$ python step5_sample_pairs_cal.py
$ python step6_sample_pairs_val.py
$ python step7_sample_pairs_dev.py

2. Train the DML, BFS and UAL layers on the training set. All results are stored in the folder results_adhominem/

   $ cd training_adhominem
$ python train_adhominem.py

3. Train O2D2 on the calibration set. All results are stored in the folder results_o2d2/

   $ cd training_o2d2
$ python train_o2d2.py

4. Perform Inference on the validation set. All results (including LEVs, predictons, attentions etc.) are stored in the folder results_o2d2/

   $ cd inference
$ python run inference.py

5. Create HTML file with attention plots for a single document. The html files will be stored in the folder results_o2d2/

   $ cd inference
$ python plot_att_html.py

6. Plot reliability diagrams, posterior histograms, or attentions. The plots will be stored in the folder results_o2d2/

   $ python plot_rel_diag.py
$ python plot_cal_hist.py
$ python plot_att_html.py

Cite the paper

If you use our code or data, please cite the paper using the following BibTeX entry:

@InProceedings{boenninghoff:2021,
  author = {Benedikt Boenninghoff and {Robert M.} Nickel and Dorothea Kolossa},
  booktitle = {{CLEF 2021 Labs and Workshops, Notebook Papers}},
  crossref = {pan:2021},
  editor = {Guglielmo Faggioli and Nicola Ferro and Alexis Joly and Maria Maistro and Florina Piroi},
  month = sep,
  publisher = {CEUR-WS.org},
  title = {{O2D2: Out-Of-Distribution Detector to Capture Undecidable Trials in Authorship Verification---Notebook for PAN at CLEF 2021}},
  url = {http://ceur-ws.org/Vol-2936/paper-158.pdf},
  year = 2021
}