Current conversational AI systems based on large language models (LLMs) are known to generate unsafe responses agreeing to offensive user input or including toxic content. Previous research aimed to alleviate the toxicity by fine-tuning LLM with manually annotated safe dialogue histories. However, the dependency on additional tuning requires substantial costs. To remove the dependency, we propose GrounDial, where response safety is achieved by grounding responses to commonsense social rules without requiring fine-tuning. A hybrid approach of in-context learning and human-norm-guided decoding of GrounDial enables the response to be quantitatively and qualitatively safer even without additional data or tuning.

Text classification is an important problem with a wide range of applications in NLP. However, naturally occurring data is imbalanced which can induce biases when training classification models. In this work, we introduce a novel contrastive learning (CL) approach to help with imbalanced text classification task. CL has an inherent structure which pushes similar data closer in embedding space and vice versa using data samples anchors. However, in traditional CL methods text embeddings are used as anchors, which are scattered over the embedding space. We propose a CL approach which learns key anchors in the form of label embeddings and uses them as anchors. This allows our approach to bring the embeddings closer to their labels in the embedding space and divide the embedding space between labels in a fairer manner. We also introduce a novel method to improve the interpretability of our approach in a multi-class classification scenario. This approach learns the inter-class relationships during training which provide insight into the model decisions. Since our approach is focused on dividing the embedding space between different labels we also experiment with hyperbolic embeddings since they have been proven successful in embedding hierarchical information. Our proposed method outperforms several state-of-the-art baselines by an average 11% F1. Our interpretable approach highlights key data relationships and our experiments with hyperbolic embeddings give us important insights for future investigations. We will release the implementation of our approach with the publication.

Jointly fine-tuning a Pre-trained Language Model (PLM) on a pre-defined set of tasks with in-context instructions has been proven to improve its generalization performance, allowing us to build a universal language model that can be deployed across task boundaries. In this work, we explore for the first time whether this attractive property of in-context instruction learning can be extended to a scenario in which tasks are fed to the target PLM in a sequential manner. The primary objective of so-called lifelong in-context instruction learning is to improve the target PLM’s instance- and task-level generalization performance as it observes more tasks. DynaInst, the proposed method to lifelong in-context instruction learning, achieves noticeable improvements in both types of generalization, nearly reaching the upper bound performance obtained through joint training.