Language models: new methods for conversational agents
Author
Prof. Thomas Hofmann
ETH Zurich
This project made several theoretical advances in the field of language models, in particular for conversational agents or dialogue systems used to answer queries.
The project achieved results in four areas that are central and foundational to the rapidly developing areas of natural language processing and conversational agents.
First, in entity detection and linking, the project contributed a novel entity-linking system that combined advanced entity embeddings, a neural attention mechanism over local context windows, and differentiable joint disambiguation inferencing. Notably, the system combines entity detection and linking. In addition, work in this area has spurred innovative follow-on work.
Second, in the area of language generation models that are foundational to conversational systems, results were obtained that address different limitations caused by biases and teacher forcing when training unconditional language models.
Third, in relation to the use of deep neural networks for generative models, the project achieved advances in learning algorithms for, and evaluation of, generative adversarial networks (GAN). It was hoped that GANs could be used for text production and conversational exchanges, which remains a challenge.
Fourth, the project achieved results related to the design of reinforcement learning agents in the setting of text-based games. Focus was on how to contend with the compositional and combinatorial nature of language that make it hard to optimise policies, and an agent was designed that was able to perform well across a family of games, rather than in only a single game.
Overall, the project made influential contributions to machine learning methodology, most notably in the areas of geometric embeddings and generative models. The results are documented in a dozen papers that encompass already highly cited papers in top conferences in Machine Learning or AI, such as NeurIPS, ICML, and AISTATS.