patrickraoulphilipp / nerd-expertprocess

Expert Processes for Named Entity Recognition and Disambiguation

Implementation of the online hedge algorithm for NERD expert processes of Reinforcement Learning for Multi-Step Expert Advice using so-called meta-dependencies as features, which are neighorhood-based assessment of available expert services. Note that the expert services for NER and NED (e.g. AIDA, DBpedia Spotlight) are not accessible anymore, but I provide a "stash" of >300K service evaluations for the integrated datasets. I also added exemplary textual features based on OpenAI's ClIP model, which were not included in the original implementation.

Instructions

1. Clone this repo.

git clone https://github.com/patrickraoulphilipp/nerd-expertprocess
cd nerd-expertprocess

2. (optional) Create a virtualenv. The implementation has been tested for Python 3.9.

virtualenv venv
source venv/bin/activate

3. Install all dependencies. You need CLIP, which will be automatically installed from the respective git repo.

pip install -r requirements.txt .

4. Download all nltk dependencies, which you can do via the python script in the scripts folder.

python scripts/install_nltk.py

5. (Optional but recommended) Download the zipped expert service stash from the following link.

Direct link: https://drive.google.com/file/d/1T94xTkOrm3gyJEB2FvK4PJnT8XqWykqC/view?usp=sharing

6. Set parameter STASH_PATH in nerd_expertprocess/ep_config.py, which should either point to an empty folder to gather the expert service results or to the downloaded & unzipped stash folder.

STASH_PATH = '/PATH/TO/FOLDER/'
...

7. Run main.py to start the search process.

python main.py

Cite as

@inproceedings{philipp2017,
 author    = {Patrick Philipp and Achim Rettinger},
 title     = {Reinforcement Learning for Multi-Step Expert Advice},
 booktitle = {AAMAS},
 year      = {2017},
 pages     = {962--971}