d2klab / jde-predict

Application that predicts business events, extract named entities and topics from news articles

jde-predict

Docker

docker-compose up

API

Prerequisites

1. Python >= 3.9

Development

python app.py

Production

python -m gunicorn -w 1 app:app

Frontend

Prerequisites

1. Node.js >= 10

Development

First, run the development server:

cd frontend/
npm ci
npm run dev

Open http://localhost:3000 with your browser to see the result.

Production

npm ci
npm start

jde-predict's Issues

De-anonymize the algorithms being used for predicting Business Events

In the demo UI (https://jde-predict.tools.eurecom.fr/), replace:

• 'Algorithme 1' --> 'BERT-based'
• 'Algorithme 2a' --> 'GPT 3.5'
• 'Algorithme 2b' --> 'Claude 1.0'
• 'Algorithme 3' -->'ZeSTE NLI'

Evaluate the quality of the annotations on news articles coming from another source than JDE

In the ISWC 2023 demo paper, we have evaluated the results of the predictions on articles taken from the same source (JDE) than the one being used for training the supervised model for predicting business events.

Let's select 5 articles from Les Echos, Le Figaro and Le Monde, which are related to financial/business and run our predictions: business events (4 algorithms), themes and named entities. Let's evaluate a posteriori the results. A markdown file detailing the articles chosen, results obtained and analysis should be produced

Design and Populate a Knowledge Graph storing the annotations of articles

The demonstration at https://jde-predict.tools.eurecom.fr/ enables to submit any article from the JDE and to visualize a number of annotations on this article namely:

• a prediction of the Business Events relevant for the given article (among 11 possible classes); 4 different algorithms are providing predictions together with a score.
• a list of named entities extracted in the given article; 3 NER tools are used (spaCy, Flair, and a pre-trained CamemBERT model) and the final results include majority voting and other post-processing.
• a prediction of the general themes (among 10 possible classes)

The goal is to materialize these annotations in a KG. The tasks are:

• Design a lightweight model to represent/identify the news article and these annotations. The Business Events could be represented as skos:Concept in a dedicated ConceptScheme. Similarly, the Themes could also be represented as skos:Concept in another ConceptScheme. Named Entity annotations could re-use the NIF ontology.
• Implement a converter for transforming the current JSON format in RDF to populate this KG
• Propose in a README a number of useful SPARQL queries for this KG