A New Broad NLP Training from Speech to Knowledge

Maxime Amblard, Miguel CouceiroNAACL 2021

This paper details the creation and structure of a novel Master's program in Natural Language Processing (NLP) at the University of Lorraine, France. It highlights the program's unique approach to integrating speech and knowledge processing alongside traditional NLP, aiming to produce well-rounded professionals for both industry and academia.

TL;DR

This paper presents a new Master's program in NLP that broadens the scope beyond traditional text processing to include speech and knowledge representation. The program's innovative curriculum and pedagogical approach are designed to equip students with a comprehensive skill set for the evolving field of language data processing.

What Problem Does This Paper Address?

The authors identify a gap in existing NLP training programs, which often remain narrowly focused on text-based tasks. They argue that the increasing complexity and interdisciplinary nature of NLP, driven by advancements in AI and deep learning, necessitate a more holistic approach. This new Master's program aims to address this by integrating speech processing and knowledge representation, preparing graduates for a wider range of industry and research challenges. The program also seeks to bridge the divide between theoretical computer science and practical linguistic applications, catering to students with diverse backgrounds.

Methods

The creation of the Master's program involved a rigorous, bottom-up methodology. A project manager was appointed to lead a working group comprising members from the LORIA and ATILF research teams. This group, which grew from 13 to 25 participants, met regularly to define the program's content and structure. The approach emphasized starting from scratch to renovate existing curricula and integrate new perspectives. A shared online workspace facilitated collaboration and documentation. The program's architecture is designed to be regular and legible, with a focus on building a common foundation in the first year and specializing in NLP applications and internships in the second. Teaching practices are renewed through project-based learning, encouraging group work and research discovery.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Broadening the NLP Scope: The program's core innovation is its explicit inclusion of speech and knowledge processing alongside traditional NLP, creating a more comprehensive training in language data processing.
  • Interdisciplinary Integration: The program successfully integrates students from diverse backgrounds (computer science, linguistics, and mixed profiles), fostering a collaborative learning environment.
  • Project-Based Learning: A strong emphasis on "supervised projects" and applied group projects allows students to gain practical experience, develop research skills, and master scientific presentation.
  • Bridging Academia and Industry: The curriculum is designed to prepare students for both research careers and industry roles, addressing the growing demand for NLP professionals in the job market.
  • French Diploma in NLP: The program offers a unique French diploma in NLP, contributing to the national landscape of specialized AI and language technology education.
  • Adaptable Teaching: The program acknowledges the need for refresher courses and differentiated instruction, particularly for technical subjects like Python programming, to accommodate varying student levels.

Key Results

The paper presents data on student admissions and profiles over six years, indicating a consistent demand for the program. The distribution of students across computer science, mixed, and linguistics profiles has been relatively balanced, facilitating the formation of effective work groups. The program's opening as a Master's has significantly increased the average number of students, with a notable rise in applicants for the M1 year. The authors observe that students who complete the full two-year program tend to achieve better results, suggesting the training's unique characteristics contribute to their success. Furthermore, a significant portion of graduates from the first cohort are pursuing PhDs, demonstrating the program's effectiveness in preparing students for academic research.

Limitations and Future Work

The authors acknowledge that transforming students with no prior programming experience into operational computer scientists, or vice-versa, within two years is challenging. They also note that the program's current structure does not allow for multiple specialized pathways, requiring students to build their profiles collaboratively. Future work includes consolidating the training by slightly increasing student intake, particularly from European countries, and developing research opportunities to offer high-quality training towards PhDs. The program is also preparing for a general evolution proposal for 2023, aiming for greater autonomy.

Summary updated

BibTeX
@inproceedings{amblard2021new,
  author = {Maxime Amblard and Miguel Couceiro},
  title = {{A New Broad NLP Training from Speech to Knowledge}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
}
Figure