For the second year running, the reciTAL research team has had one of its scientific publications accepted at the NeurIPS conference. The article “To Beam, or Not to Beam: That is a Question of Cooperation for Language GANs” focuses on a new method for training text generation models.

NeurIPS: what's it all about?

NeurIPS (Conference on Neural Information Processing Systems) is one of the most important scientific events in artificial intelligence. Every year, NeurIPS brings together thousands of researchers from all over the world to present their work and discuss the future of machine learning. Members of the reciTAL R&D team will therefore be among the speakers at this event, to be held from December 6 to 14, 2021.

The new NeurIPS publication is the culmination of several years’ research. This is a great source of pride and recognition for the reciTAL R&D team. The 6 publications accepted in 2021 (
5 at EMNLP
and one at NeurIPS) testify not only to the quality of the research carried out by the reciTAL team, but also to our commitment to advancing research in Artificial Intelligence and Machine Learning. “says Thomas Scialom, Research Scientist at reciTAL.

About the accepted publication

The paper accepted at NeurIPS focuses on Generative Models.

Generative models are becoming the norm in NLP. For example, the largest language model, GPT-3, considers all language tasks as generative tasks, such as classification, information extraction or Question Answering.

Therefore, improving generative models can potentially improve all tasks in Language Processing.

Until now, generative models have tended to generate texts that are easy to discern from a human-written text, and contain many errors.

One way around this problem is to use GANs (Generative Adversarial Networks). In this case, a discriminator tries to detect whether a text is generated by a machine or written by a human, and the machine trains itself to fool the discriminator.

With this learning process, the quality of the text generated by the model is greatly improved.

Partager sur