As Large Language Models (LLMs) become increasingly prevalent, their generated outputs are proliferating across the web, risking a future where machine-generated content dilutes human-authored text. Since online data is the primary resource for LLM pre-training, subsequent models could be trained on an unknown portion of synthetic samples. This could lead to model collapse, a degenerative process whereby LLMs reinforce their own errors, reduce output diversity, and ultimately yield declining performance. In this study, we investigate the impact of decoding strategy on model collapse, analysing the text characteristics at each model generation, the similarity to human references, and the resulting model performance. Using the decoding strategies that lead to the most significant degradation, we evaluate model collapse in a more realistic scenario where the origin of the data (human or synthetic) is unknown. We train a machine-generated text detector and propose an importance resampling approach to prevent model collapse by up-sampling likely human content in the training data. Our method is validated on four LLMs from two model families (GPT-2 and SmolLM2), across a range of model sizes (124M to 1.7B). We demonstrate that it not only prevents model collapse but also improves performance compared to training on purely human data, underscoring the benefit of synthetic samples and the importance of data curation.

Identifying named entities such as a person, location or organization, in documents can highlight key information to readers. Training Named Entity Recognition (NER) models requires an annotated data set, which can be a time-consuming labour-intensive task. Nevertheless, there are publicly available NER data sets for general English. Recently there has been interest in developing NER for legal text. However, prior work and experimental results reported here indicate that there is a significant degradation in performance when NER methods trained on a general English data set are applied to legal text. We describe a publicly available legal NER data set, called E-NER, based on legal company filings available from the US Securities and Exchange Commission’s EDGAR data set. Training a number of different NER algorithms on the general English CoNLL-2003 corpus but testing on our test collection confirmed significant degradations in accuracy, as measured by the F1-score, of between 29.4% and 60.4%, compared to training and testing on the E-NER collection.

Previous research has linked psychological and social variables to physical health. At the same time, psychological and social variables have been successfully predicted from the language used by individuals in social media. In this paper, we conduct an initial exploratory study linking these two areas. Using the social media platform of Twitter, we identify users self-reporting symptoms that are descriptive of influenza-like illness (ILI). We analyze the tweets of those users in the periods before, during, and after the reported symptoms, exploring emotional, cognitive, and structural components of language. We observe a post-ILI increase in social activity and cognitive processes, possibly supporting previous offline findings linking more active social activities and stronger cognitive coping skills to a better immune status.