Event Detection & Extraction

Repository for QA-based event detection and extraction from news using a fine-tuned roberta-large model.

Installation

1. Clone this repository:

https://github.com/sodalabsio/event-detection-extraction.git

2. Create a virtual environment (optional but recommended) and activate it:

virtualenv -p python3 venv
source venv/bin/activate

3. Install the dependencies:

cd event-detection-extraction
pip install -r requirements.txt

NB: Tested on Python 3.8.

Running the code

Here's the usage in detail (also accessible through python run.py -h):

usage: run.py [-h] --query QUERY --start-date START_DATE --end-date END_DATE --source {news,twitter} --output-path OUTPUT_PATH [--trusted-only TRUSTED_ONLY] [--event-confidence EVENT_CONFIDENCE]
              [--top-k TOP_K]

optional arguments:
  -h, --help            show this help message and exit
  --query QUERY         Query to search for. Country or place full name, for example
  --start-date START_DATE
                        Start date of the event in YYYY-MM-DD format
  --end-date END_DATE   End date of the event in YYYY-MM-DD format
  --source {news,twitter}
                        Source of events to pull from such as news or twitter (only news is supported currently)
  --output-path OUTPUT_PATH
                        Output directory to write the results
  --trusted-only TRUSTED_ONLY
                        Whether to use trusted news sources to filter by (optional)
  --event-confidence EVENT_CONFIDENCE
                        A minimum confidence threshold for events
  --top-k TOP_K         The number of top events to surface based on frequency (should be less than 15)

For example:

    python3 run.py  \
    --query=Ukraine  \
    --start-date=2022-09-06  \
    --end-date=2022-09-30  \
    --source=news  \
    --output-path=./output  \
    --trusted-only=True  \
    --event-confidence=0.5  \
    --top-k=10

Note

• The query string is case-insensitive i.e. "Ukraine" and "ukraine" are equivalent
• The news and event detection parts of the pipeline are cached once run (for a given --query, --start-date and --end-date) to avoid re-running them. To re-run the pipeline, delete the cache files in the --output-path directory
• The pipeline can be subsequently run with different values for --event-confidence and --top-k to generate different event streamgraphs
• All data and graphs are exported under the --output-path directory. The directory is automatically created during execution
• To add/remove trusted news sources (as used when --trusted-only=True), edit here
• Google News returns max 100 news items

References

• Lyu, Q., Zhang, H., Sulem, E., & Roth, D. (2021, August). Zero-shot event extraction via transfer learning: Challenges and insights. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers) (pp. 322-332).
• Hamborg, F., Breitinger, C., & Gipp, B. (2019). Giveme5w1h: A universal system for extracting main events from news articles. arXiv preprint arXiv:1909.02766.