1

We show that bias can also be analyzed by using a large collection of individuals' names and there are certain advantages to this method compared to traditional methods that use referring phrases.

We show that language models are biased towards Western moral justifications. Interestingly we also find that LLMs can be overaligned towards religious concepts.

We show diverse data augmentations can improve downstream language model performance.

We show the inconsistency in measuring social biases in language models using text completion, highlighting how different methods and settings affect results also offer guidelines for more accurate bias reporting and provides the study’s code.

We show T0, a text-to-text language model, exhibits more social bias in question-answer formats than in premise-hypothesis formats, highlighting the impact of input forms on bias also include a new bias benchmark and provides code and data.

We show that a multitask feature extraction and aggregation can model human perception of similarity of translated sentences much better than standart n-gram level distance metrics.

We show that a multitask feature extraction and aggregation can model human perception of similarity of translated sentences much better than standart n-gram level distance metrics.

New method for assessing the quaility of similarity evaluators and showing potential of Transformer-based language models in replacing BLEU and ROUGE.