Projects
topicx: Clustering with Contextual Embeddings for Topics
A framework for running topic models via clustering contextual embeddings. By clustering embeddings obtained from pre-trained language models (e.g., BERT) along with a word-selecting method, we can obtain coherent and diverse topics, which are simpler and more efficient than neural topic models.
Stay Moral and Explore: Learn to Behave Morally in Text-based Games
A general algorithm is named Moral Awareness Adaptive Learning (MorAL). It boosts an agent’s moral capacity. This is achieved using a moral-aware learning model plugin. The algorithm lets the agent learn tasks and morality adaptively.
Self-imitation Learning for Action Generation in Text-based Games
A confidence-based self-imitation model generates action candidates for the RL agent in text-based games. It utilizes self-imitation learning to rank and leverage valuable past trajectories. Also, a confidence-based strategy is used to assess the LM’s confidence relative to a state.