Foundation models pre-trained on large corpora demonstrate significant gains across many natural language processing tasks and domains e.g., law, healthcare, education, etc. However, only limited efforts have investigated the opportunities and limitations of applying these powerful models to science and security applications. In this work, we develop foundation models of scientific knowledge for chemistry to augment scientists with the advanced ability to perceive and reason at scale previously unimagined. Specifically, we build large-scale (1.47B parameter) general-purpose models for chemistry that can be effectively used to perform a wide range of in-domain and out-of-domain tasks. Evaluating these models in a zero-shot setting, we analyze the effect of model and data scaling, knowledge depth, and temporality on model performance in context of model training efficiency. Our novel findings demonstrate that (1) model size significantly contributes to the task performance when evaluated in a zero-shot setting; (2) data quality (aka diversity) affects model performance more than data quantity; (3) similarly, unlike previous work, temporal order of the documents in the corpus boosts model performance only for specific tasks, e.g., SciQ; and (4) models pre-trained from scratch perform better on in-domain tasks than those tuned from general-purpose models like Open AI’s GPT-2.

With the increasing use of machine-learning driven algorithmic judgements, it is critical to develop models that are robust to evolving or manipulated inputs. We propose an extensive analysis of model robustness against linguistic variation in the setting of deceptive news detection, an important task in the context of misinformation spread online. We consider two prediction tasks and compare three state-of-the-art embeddings to highlight consistent trends in model performance, high confidence misclassifications, and high impact failures. By measuring the effectiveness of adversarial defense strategies and evaluating model susceptibility to adversarial attacks using character- and word-perturbed text, we find that character or mixed ensemble models are the most effective defenses and that character perturbation-based attack tactics are more successful.

Evaluating model robustness is critical when developing trustworthy models not only to gain deeper understanding of model behavior, strengths, and weaknesses, but also to develop future models that are generalizable and robust across expected environments a model may encounter in deployment. In this paper, we present a framework for measuring model robustness for an important but difficult text classification task – deceptive news detection. We evaluate model robustness to out-of-domain data, modality-specific features, and languages other than English. Our investigation focuses on three type of models: LSTM models trained on multiple datasets (Cross-Domain), several fusion LSTM models trained with images and text and evaluated with three state-of-the-art embeddings, BERT ELMo, and GloVe (Cross-Modality), and character-level CNN models trained on multiple languages (Cross-Language). Our analyses reveal a significant drop in performance when testing neural models on out-of-domain data and non-English languages that may be mitigated using diverse training data. We find that with additional image content as input, ELMo embeddings yield significantly fewer errors compared to BERT or GLoVe. Most importantly, this work not only carefully analyzes deception model robustness but also provides a framework of these analyses that can be applied to new models or extended datasets in the future.