Skip to content

Latest commit

 

History

History
80 lines (58 loc) · 3.8 KB

README.md

File metadata and controls

80 lines (58 loc) · 3.8 KB
Raw

Qualle

Implementation of Qualle Framework as proposed in the paper Content-Based Quality Estimation for Automatic Subject Indexing of Short Texts under Precision and Recall Constraints and accompanying source code by Martin Toepfer and Christin Seifert.

The framework allows to train a model which can be used to predict the quality of the result of the application of a multi-label classification (MLC) method on a document. Currently, only the Recall is predicted for a document, but in principle any document-level quality estimation (like prediction of precision) can be implemented in the future.

Qualle provides a Command-line interface to train and evaluate models. Beyond that, a REST Webservice for predicting the Recall of a MLC result is provided.

Command line interface (CLI)

In order to run the CLI, you must install the packages from the Pipfile. The interface is than accessible from the module qualle.main. To see the help message, run (inside the Qualle directory)

python -m qualle.main -h

Train

You must provide a train data file in order to train a model. This file is a tabular-separated file (tsv) in the format (tabular is marked with \t)

document-content\tpredicted_labels_with_scores\ttrue_labels

where

  • document-content is a string describing the content of the document (more precise: the string on which the MLC method is trained on), e.g. the title
  • predicted_labels_with_scores is a comma-separated list of pairs predicted_label:confidence-score This is in principle the output of the MLC method.
  • true_labels is a comma-separated list of true labels (ground truth)

For example, a row in the data file could look like

Optimal investment policy of the regulated firm\tConcept0:0.5,Concept1:1\tConcept0,Concept3

To train a model, use the main module inside qualle, e.g.:

python -m qualle.main train /path/to/train_data_file /path/to/output/model

It is also possible to use Label Calibration using the Subthesauri of a Thesaurus (like STW) as categories (please read the paper for more explanation). Consult the help for the required options.

Evaluate

You must provide a test data file and the path to a trained model in order to evaluate a model. The test data file has the same format as the train data file. Metrics like the Explained Variation are printed out, describing the quality of the Recall Prediction (please consult the paper for more information).

REST interface

To perform the prediction on a MLC result, a REST interface can be started. uvicorn is used as HTTP Server. You can also use any ASGI Server implementation and create the ASGI app directly with the method qualle.interface.rest.create_app. You need to provide the environment variable MODEL_FILE with the path to the model (see qualle.interface.config.RESTSettings).

The REST Endpoint expects a HTTP POST with the result of a MLC for a list of documents as body. The format is JSON as specified in qualle/openapi.json. You can also use the Swagger UI accessible at http://address_of_server/docs to play a bit around.

Deployment with Docker

You can use the Dockerfile included in this project to build a Docker Image. E.g.:

docker build -t qualle .

Per default, gunicorn is used to run the REST interface on 0.0.0.0:8000 You need to pass the required settings per environment variable. E.g.:

docker run --rm -it --env model_file=/model -v /path/to/model:/model -p 8000:8000 qualle-test

Of course you can also use the docker image to train or evaluate by using a different command as input to docker run.