epfl-dlab/pti-candgen

Source code for "Efficient Entity Candidate Generation for Low-Resource Languages" Source code for "Entity-for-mention Candidate Generation for Low-resource Languages" # Data Download the data from this link It contains 18 folders (one per language) which should be put inside the mentions_dumps folder. The files `interlanguage_links_wikidata-20190901.csv` and `qid2title_CHAR.pkl` have to be in the same folder as the README file. # Virtual Environment Setup the virtual environment named `pti` to install all the required dependencies `conda env create -f pti.yml`. Activate the installed environment `conda activate pti` # Running PTI and Charagram PTI and Charagram can be run by simply executing `python PTI/main.py alpha lambda target_lang pivot_lang` and `python CHARAGRAM/main.py mu amount_training_data target_lang pivot_lang` respectively. Zero-shot setting is enabled by setting `alpha=-1` or `mu=-1`. Following guidelines by the authors of Charagram, `amount_training_data` is set to 80,000 for the main experiments contained in the paper. Both the pivot and target language are represented with the two or three character code indicated in the submission. # Running WikiPriors To run WikiPriors simply execute `python WikiPriors/main.py args`. where the arguments `args` are listed below: - --tlang: target language. - --plang: pivot language. - --ncands: number of retrieved candidate entities. The default value is 30. - --zeroshot: boolean to indicate whether the learning setting is zero-shot or not.