Knowledge graph models manipulator which allows to perform various evaluations of provided embedders. Currently there is only an incomplete support for fetching wikidata triples and saving them in the format acceptable for the OpenKE toolkit (see examples of the generated datasets in the folder Assets/Demo).

For being able to run python-based code you need to create a new environment and install the project dependencies:

conda env create --file environment.yml

Currently on hyperparameter search in supported, which can be initiated using the following command which configures 2 concurrent workers to process cross-validation splits for a given dataset and Assets/Log/default.txt as a log destination file path:

swift run relentness hsearch -e grapex-latest 17 19 -n 2 -l default

To generate a sample from wikidata using default params, execute the following command:

swift build && ./.build/debug/relentness Assets/Demo

The generated files will be saved in the directory Assets/Demo. To list all available option for the sample subcommand, please, use the following call:

./.build/debug/relentness sample --help

To test a model on a given dataset with transe model:

python -m relentness test ./Assets/Corpora/Demo/0000 -m transe

Aside from transe there is a complex model supported as well at the moment which requires a similar call.

The applications allows to export results of models comparison into google spreadsheet with nice formatting which automatically highlights all the main points.

It is possible to interact with the application via telegram bot. Upon startup you are able to provide any access token and fine-tune the app to send a wide range of notifications, the most interesting of which is the message that informs you about the need of updating OAuth2 access token to a third-party service. On the screen below an example of a short chat with telegram bot is provided.