Towards Abstractive Summarization of DBpedia Abstracts Using Language Models

This repository contains code for our paper "Abstractive Summarization of DBpedia Abstracts Using Language Models." We propose an approach using pre-trained language models, specifically BART and T5, to generate short and comprehensive summaries for DBpedia abstracts in six languages (English, German, French, Italian, Spanish, and Dutch).

The pipeline of DBpedia summarization using language models

Dependencies

• bert-score 0.3.12
• ipykernel 6.17.1
• ipython 8.6.0
• nltk 3.7
• notebook 6.5.2
• pandas 1.5.1
• spacy 3.4.3
• torch 1.13.0
• transformers 4.25.1

Repository Structure:

├── data
│   ├── info.md
├── data_crowd
│   ├── de_crowd.csv
│   ├── en_crowd.csv
│   ├── es_crowd.csv
│   ├── fr_crowd.csv
│   ├── it_crowd.csv
│   ├── nl_crowd.csv
├── data_eval
│   ├── de_100_summaries.csv
│   ├── en_100_summaries.csv
│   ├── es_100_summaries.csv
│   ├── fr_100_summaries.csv
│   ├── it_100_summaries.csv
│   ├── nl_100_summaries.csv
├── full_abstracts
│   └── info.md
├── short_abstracts
│   └── info.md
├── baselines.ipynb
├── data_creation.ipynb
├── dbepdia-summarization.png
├── DBpepdia-abstractive-summarization.md
├── LICENSE
├── README.md
├── requirements.txt
├── summarization-cuda1.py
└── summarization-split.py

Installation

To install the dependencies, run:

pip install -r requirements.txt

Data Preparation

1. Download the DBpedia abstract file (in .ttl format) for the desired language from this source and place it in the full_abstracts folder.
2. Download the DBpedia short abstract file (in .ttl format) for the desired language from this source and unzip it in the short_abstracts folder.
3. Run the data_creation notebook. The final dataframes should be located in the data folder.

Crowdsourcing evaluation

The complete result of evaluation made by crowdsourcing agents are presented in data_crowd folder. The data were divided by language. Each file contains column 'choice' which demonstrates which summary were chosen by crowdworker.

Pretrained Models

1. The data for generating summaries is located in the data folder in .csv files.
2. The baselines.ipynb notebook contains the code for running the pretrained models (T5, BART, and BART-CNN).
3. The original and short abstracts and generated summaries used for crowdsourcing evaluation are stored in in data_eval folder.

Generating Astractive Summaries for The Full DBpedia Abstracts

More details are descripted here about downloading and processing the entire DBpedia abstracts.

---

How to cite

TBD

Contact

TBD