NLP architectures in the age of end-to-end deep-learning systems
Bringing together both worlds and discuss the future of NLP engineering
Tuesday, May 12, 2020
Co-located with LREC2020, Marseille
In the past, almost all NLP tasks were tackled with a pipeline approach combining within each pipeline different systems solving clearly defined tasks (e.g. POS tagging, parsing, semantic role labeling etc.). As the different implementations often were not directly compatible several NLP frameworks were introduced (e.g. GATE, DKPro Core, NLTK, spaCy, etc.) that provided the “glue” to hold everything together.
Recently, the trend has clearly focused on end-to-end deep learning systems that directly solve the task at hand without explicitly representing intermediate, linguistically motivated levels. Of course this view is an exaggeration, as the two worlds are still interconnected. Recently, on the one hand, NLP pipeline elements have been discovered in end-to-end systems [1], on the other hand, end-to-end systems have been integrated in traditional pipeline architectures.
The workshop is motivated by the perceived growing divide between classical pipeline architectures and the recent focus on end-to-end approaches. We believe that there is more common ground than currently considered on both sides. It would be beneficial to bring all interested researchers together at this workshop to explore future directions.
[1] Tenney, I., Das, D., Pavlick, E. 2019. BERT Rediscovers the Classical NLP Pipeline. arXiv e-prints arXiv:1905.05950
Important Dates:
- Submission deadline : February 14, 2020
- Author notification: March 13, 2020
- Camera ready: April 2, 2020
- Workshop: May 12, 2020
Note: All deadlines are 11:59PM UTC-12:00 ("anywhere on Earth")
Organizing committee:
- Piush Aggarwal, Language Technology Lab, University of Duisburg-Essen, Germany
- Mohamed Karim Bouzoubaa, ALELM Lab, University Mohammed V in Rabat, Morocco
- Richard Eckart de Castilho, Ubiquitous Knowledge Processing Lab, Technische Universität Darmstadt, Germany
- Torsten Zesch, Language Technology Lab, University of Duisburg-Essen, Germany
Contact e-mail address: torsten.zesch@uni-due.de