Joint event CORLI LIFT TAL – May, 15-16, Collaborative corpus annotation

Date/heure
Date(s) - 15/05/2023 - 16/05/2023
14h00 - 12h30

Catégories

The CORLI consortium and the LIFT and TAL research groups are hosting an event surrounding Amir Zeldes’ conference on Monday, May 15 from 2 to 5:30 pm at Paris Cité University (Campus des grands moulins, Olympe de Gouge building).

PROGRAM

Monday, May 15 – 2:00 pm to 5:00 pm

Where? Room 720, Olympe de Gouges building

Presentation of the CORLI consortium and LIFT and TAL research groups
Conference by Amir Zeldes followed by a round table on corpus annotation and annotated corpora.

Crafting Rich and Diverse Data for Computational Models of Discourse

Thanks to pre-trained language models, computational tools are now better than ever at extracting entities, events, and other information from arbitrary text and speech. However, models are only successful and useful when they learn from the right kind of data, making the representations and materials we choose, and what we teach practitioners in the field, more important than ever. In this talk I will present the Georgetown University Multilayer (GUM) family of datasets, which are constructed as part of the Computational Linguistics curriculum at Georgetown, and target discourse level representations of spoken and written language across 20 distinct text types. After discussing tools, architectures and pedagogical challenges for developing data in an academic curriculum, I will present a series of studies illustrating the added value of richly annotated resources, the importance of genre diversity, limitations of current models and some new theoretical directions in computational discourse models. Our results show that carefully crafted datasets are crucial for a range of popular tasks, including salient entity recognition, Wikification, coreference resolution, automatic summarization, discourse parsing, and more.

Tuesday, May 16 – 10:00 am to 12:30 pm

Where? Room 531, Olympe de Gouges building

CORLI-GUM project meeting

Information

Flyer (pdf)

Loading...

Taking too long?