Large language models are meticulously aligned to be both helpful and harmless. However, recent research points to a potential overkill which means models may refuse to answer benign queries. In this paper, we investigate the factors for overkill by exploring how models handle and determine the safety of queries. Our findings reveal the presence of shortcuts within models, leading to an over-attention of harmful words like 'kill' and prompts emphasizing safety will exacerbate overkill. Based on these insights, we introduce Self-Contrastive Decoding (Self-CD), a training-free and model-agnostic strategy, to alleviate this phenomenon. We first extract such over-attention by amplifying the difference in the model's output distributions when responding to system prompts that either include or omit an emphasis on safety. Then we determine the final next-token predictions by downplaying the over-attention from the model via contrastive decoding. Empirical results indicate that our method has achieved an average reduction of the refusal rate by 20\% while having almost no impact on safety.

We describe the systems of the University of Alberta team for the SemEval-2023 Visual Word Sense Disambiguation (V-WSD) Task. We present a novel algorithm that leverages glosses retrieved from BabelNet, in combination with text and image encoders. Furthermore, we compare language-specific encoders against the application of English encoders to translated texts. As the contexts given in the task datasets are extremely short, we also experiment with augmenting these contexts with descriptions generated by a language model. This yields substantial improvements in accuracy. We describe and evaluate additional V-WSD methods which use image generation and text-conditioned image segmentation. Overall, the results of our official submission rank us 18 out of 56 teams. Some of our unofficial results are even better than the official ones. Our code is publicly available at https://github.com/UAlberta-NLP/v-wsd.

Dependency parsing is a crucial step towards deep language understanding and, therefore, widely demanded by numerous Natural Language Processing applications. In particular, left-to-right and top-down transition-based algorithms that rely on Pointer Networks are among the most accurate approaches for performing dependency parsing. Additionally, it has been observed for the top-down algorithm that Pointer Networks' sequential decoding can be improved by implementing a hierarchical variant, more adequate to model dependency structures. Considering all this, we develop a bottom-up-oriented Hierarchical Pointer Network for the left-to-right parser and propose two novel transition-based alternatives: an approach that parses a sentence in right-to-left order and a variant that does it from the outside in. We empirically test the proposed neural architecture with the different algorithms on a wide variety of languages, outperforming the original approach in practically all of them and setting new state-of-the-art results on the English and Chinese Penn Treebanks for non-contextualized and BERT-based embeddings.