This project contains the source code for my bachelor's thesis "Abstractive Text Summarization of Meetings".
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.
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.
python main.py --run_mode=train_and_evaluate
starts the training.
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.
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.
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.
The AMI Corpus license can be found here: AMI Meeting Corpus License.
Main parts of the code are taken from the Texar examples for BERT and Transformers. They can be found under the following links:
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.
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.
This project is licensed under the Apache License 2.0.