Abstractive Summarization of Meetings

This project contains the source code for my bachelor's thesis "Abstractive Text Summarization of Meetings".

Requirements

This project was only tested with Python 3.6 but should also work with more recent version of Python. For dependency versions, take a look at the requirements.txt file.

Execution

Preparing the data

python prepare_data.py

reads the data.[train|dev|test].tsv files and generates 3 TFRecord data files train.tf_record, eval.tf_record, and test.tf_record. These files are used for training.

Training

python main.py --run_mode=train_and_evaluate

starts the training.

Testing

python main.py --run_mode=test

can be used to calculate BLEU and ROUGE scores on the test data. It will print the results into the console and write the three files test-inputs.txt, test-predictions.txt, test-targets.txt in the /outputs folder. These files contain the sentences in a human readable format.

Predicting

python main.py --run_mode=predict

takes the content from the /data/predict.txt file and creates two files in the output-folder: predict-inputs.txt and predict-predictions.txt.

Credits

Data

The data from the predict.txt and data.[train|dev|test].tsv files is taken from the AMI Corpus and processed using the NITE XML Toolkit. The code that parses the corpus can be found at Meeting-Parser.

License

The AMI Corpus license can be found here: AMI Meeting Corpus License.

Code

Main parts of the code are taken from the Texar examples for BERT and Transformers. They can be found under the following links:

• BERT example
• Transformer example

These examples are licensed under the Apache License 2.0. Copied files contain a link to their original version in the file header. Any of my modifications are also licensed under the same license.

Inspiration

This project was inspired by the GitHub repository Abstractive Summarization With Transfer Learning. This project uses no source code of the repository, though. The repository is also based on the Texar examples and thus has similar code.

License

This project is licensed under the Apache License 2.0.