Berlin Language AI Meetup #1

PRESENTATION 1
Choose your metrics wisely: Investigating decoding criteria for NLP
by David Vilar Torres (Senior Research Scientist at Google)

Most machine learning models, including LLMs, generate their output guided by the probability distribution they define. Taking machine translation as a specific example, in this talk we will explore how this criterion is not always the best choice, and can lead to suboptimal decisions. We will discuss Minimum Bayes Risk decoding as an alternative generation method that can optimize arbitrary performance metrics at generation time. These improvements come however at the expense of a high computational cost, so we will present methods to make this type of decoding more efficient. But all that glisters is not gold, and we will finish with a cautionary tale about overoptimization.

—

PRESENTATION 2
How PONS Langenscheidt is using AI to rethink language (learning)
by Lars Janzik (CEO of PONS Langenscheidt)

PONS Langenscheidt has a long history and high brand recognition as a provider of books on translation, reference, and language learning. Although apps and digital services have been part of the portfolio for over 20 years, the emergence of machine translation and AI has changed everything. This creates great opportunities for PONS Langenscheidt and other traditional educational companies, but also new competition and risks, in parallel with changing usage and learning behavior.