Natural language is an appealing medium for explaining how large language models process and store information, but evaluating the faithfulness of such explanations is challenging. To help address this, we develop two modes of evaluation for natural language explanations that claim individual neurons represent a concept in a text input. In the *observational mode*, we evaluate claims that a neuron a activates on all and only input strings that refer to a concept picked out by the proposed explanation E. In the *intervention mode*, we construe E as a claim that neuron a is a causal mediator of the concept denoted by E. We apply our framework to the GPT-4-generated explanations of GPT-2 XL neurons of Bills et al. (2023) and show that even the most confident explanations have high error rates and little to no causal efficacy. We close the paper by critically assessing whether natural language is a good choice for explanations and whether neurons are the best level of analysis.

Timely and accurate extraction of Adverse Drug Events (ADE) from biomedical literature is paramount for public safety, but involves slow and costly manual labor. We set out to improve drug safety monitoring (pharmacovigilance, PV) through the use of Natural Language Processing (NLP). We introduce BioDEX, a large-scale resource for Biomedical adverse Drug Event eXtraction, rooted in the historical output of drug safety reporting in the U.S. BioDEX consists of 65k abstracts and 19k full-text biomedical papers with 256k associated document-level safety reports created by medical experts. The core features of these reports include the reported weight, age, and biological sex of a patient, a set of drugs taken by the patient, the drug dosages, the reactions experienced, and whether the reaction was life threatening. In this work, we consider the task of predicting the core information of the report given its originating paper. We estimate human performance to be 72.0% F1, whereas our best model achieves 59.1% F1 (62.3 validation), indicating significant headroom. We also begin to explore ways in which these models could help professional PV reviewers. Our code and data are available at https://github.com/KarelDO/BioDEX.

One of the major challenges in sign language translation from a sign language to a spoken language is the lack of parallel corpora. Recent works have achieved promising results on the RWTH-PHOENIX-Weather 2014T dataset, which consists of over eight thousand parallel sentences between German sign language and German. However, from the perspective of neural machine translation, this is still a tiny dataset. To improve the performance of models trained on small datasets, transfer learning can be used. While this has been previously applied in sign language translation for feature extraction, to the best of our knowledge, pretrained language models have not yet been investigated. We use pretrained BERT-base and mBART-50 models to initialize our sign language video to spoken language text translation model. To mitigate overfitting, we apply the frozen pretrained transformer technique: we freeze the majority of parameters during training. Using a pretrained BERT model, we outperform a baseline trained from scratch by 1 to 2 BLEU-4. Our results show that pretrained language models can be used to improve sign language translation performance and that the self-attention patterns in BERT transfer in zero-shot to the encoder and decoder of sign language translation models.