Understanding the Pipeline: Diagrams
A visual walkthrough of the extraction pipeline and the predicate-linking step.
This page is dedicated to providing weekly updates on the project.
Read the blog posts or check About for more info.
A visual walkthrough of the extraction pipeline and the predicate-linking step.
Investigating why IndIE's BenchIE score looked wrong, closing out the property audit, and publishing all models and datasets.
Establishing real baselines against IndIE and testing the pipeline on Gujarati and Rajasthani.
Being precise about how every evaluation number is actually calculated, and testing whether few-shot prompting helps.
Full-scale evaluation across the entire dataset, plus rebuilding the human review tool around real pipeline data.
Eighth week of the coding period, building a predicate-linking gold set and running the full pipeline end to end for the first time.
Seventh week of the coding period, streamlining the training pipeline and reviewing real model output closely.
Finding and fixing a scoring bias that was undervaluing real Wikipedia sentences, then getting the training pipeline running end to end.
Testing embedding models at full ontology scale and building real DBpedia ground truth for BenchIE.
Comparing embedding models for ontology alignment and building the first version of the human review tool.
Third week of the coding period, auditing the existing training dataset and building the ontology alignment layer.
Second week of the coding period, researching data requirements for fine-tuning and looking into annotation tooling.
First week of the coding period, understanding the existing DBpedia Hindi Chapter pipeline and prior GSoC work.