Large Language Models (LLMs) reproduce and exacerbate the social biases present in their training data, and resources to quantify this issue are limited. While research has attempted to identify and mitigate such biases, most efforts have been concentrated around English, lagging the rapid advancement of LLMs in multilingual settings. In this paper, we introduce a new multilingual parallel dataset SHADES to help address this issue, designed for examining culturally-specific stereotypes that may be learned by LLMs. The dataset includes stereotypes from 20 regions around the world and 16 languages, spanning multiple identity categories subject to discrimination worldwide. We demonstrate its utility in a series of exploratory evaluations for both “base” and “instruction-tuned” language models. Our results suggest that stereotypes are consistently reflected across models and languages, with some languages and models indicating much stronger stereotype biases than others.

Large language model development relies on the pre-train-then-align paradigm, in which the model is typically pre-trained on a large text corpus and undergoes a tuning stage to align the model with human preference or downstream tasks. We investigate the relationship between pre-training and supervised fine-tuning by considering multiple tasks as well as different pre-trained model checkpoints. Our results on 18 datasets and two models suggest that i) although the model benefits significantly through supervised fine-tuning, it may forget previously known domain knowledge and tasks that are not seen during fine-tuning; ii) the model exhibits high sensitivity to evaluation prompts after supervised fine-tuning, but this sensitivity can be alleviated through further pre-training; iii) continual pre-training improves the model in a latent way that manifests after fine-tuning; iv) The model can already solve some tasks after pre-training while fine-tuning most benefits datasets where the model does not show capability during pre-training.

Johns Hopkins University (JHU) submitted systems for all eight language pairs in the 2024 Low-Resource Language Track. The main effort of this work revolves around fine-tuning large and publicly available models in three proposed systems: i) end-to-end speech translation (ST) fine-tuning of Seamless4MT v2; ii) ST fine-tuning of Whisper; iii) a cascaded system involving automatic speech recognition with fine-tuned Whisper and machine translation with NLLB. On top of systems above, we conduct a comparative analysis on different training paradigms, such as intra-distillation for NLLB as well as joint training and curriculum learning for SeamlessM4T v2. Our results show that the best-performing approach differs by language pairs, but that i) fine-tuned SeamlessM4T v2 tends to perform best for source languages on which it was pre-trained, ii) multi-task training helps Whisper fine-tuning, iii) cascaded systems with Whisper and NLLB tend to outperform Whisper alone, and iv) intra-distillation helps NLLB fine-tuning.

NLP models have progressed drastically in recent years, according to numerous datasets proposed to evaluate performance. Questions remain, however, about how particular dataset design choices may impact the conclusions we draw about model capabilities. In this work, we investigate this question in the domain of compositional generalization. We examine the performance of six modeling approaches across 4 datasets, split according to 8 compositional splitting strategies, ranking models by 18 compositional generalization splits in total. Our results show that: i) the datasets, although all designed to evaluate compositional generalization, rank modeling approaches differently; ii) datasets generated by humans align better with each other than with synthetic datasets, or than the latter among themselves; iii) generally, whether datasets are sampled from the same source is more predictive of the resulting model ranking than whether they maintain the same interpretation of compositionality; and iv) specific lexical items in dataset impacts the measurement consistency. Overall, our results demonstrate that much work remains to be done when it comes to assessing whether popular evaluation datasets measure what they intend to measure, and suggests that elucidating more rigorous standards for establishing the validity of evaluation sets could benefit the field.

Generative models have been widely applied to solve extractive tasks, where parts of the input is extracted to form the desired output, and achieved significant success. For example, in extractive question answering (QA), generative models have constantly yielded state-of-the-art results. In this work, we study the issue of tokenization inconsistency that is commonly neglected in training these models. This issue damages the extractive nature of these tasks after the input and output are tokenized inconsistently by the tokenizer, and thus leads to performance drop as well as hallucination. We propose a simple yet effective fix to this issue and conduct a case study on extractive QA. We show that, with consistent tokenization, the model performs better in both in-domain and out-of-domain datasets, with a notable average of +1.7 F1 gain when a BART model is trained on SQuAD and evaluated on 8 QA datasets. Further, the model converges faster, and becomes less likely to generate out-of-context answers. Our results demonstrate the need for increased scrutiny regarding how tokenization is done in extractive tasks and the benefits of consistent tokenization during training.