Many models are pretrained on redacted text for privacy reasons. Clinical foundation models are often trained on de-identified text, which uses special syntax (masked) text in place of protected health information. Even though these models have increased in popularity, there has been little effort in understanding the effects of training them on redacted text. In this work, we pretrain several encoder-only models on a dataset that contains redacted text and a version with replaced realistic pseudo text. We then fine-tuned models for the protected health information de-identification task and show how our methods significantly outperform previous baselines. The contributions of this work include: a) our novel, and yet surprising findings with training recommendations, b) redacted text replacements used to produce the pseudo dataset, c) pretrained embeddings and fine-tuned task specific models, and d) freely available pseudo training dataset generation and model source code used in our experiments.

When a language model is trained to predict natural language sequences, its prediction at each moment depends on a representation of prior context. What kind of information about the prior context can language models retrieve? We tested whether language models could retrieve the exact words that occurred previously in a text. In our paradigm, language models (transformers and an LSTM) processed English text in which a list of nouns occurred twice. We operationalized retrieval as the reduction in surprisal from the first to the second list. We found that the transformers retrieved both the identity and ordering of nouns from the first list. Further, the transformers' retrieval was markedly enhanced when they were trained on a larger corpus and with greater model depth. Lastly, their ability to index prior tokens was dependent on learned attention patterns. In contrast, the LSTM exhibited less precise retrieval, which was limited to list-initial tokens and to short intervening texts. The LSTM's retrieval was not sensitive to the order of nouns and it improved when the list was semantically coherent. We conclude that transformers implemented something akin to a working memory system that could flexibly retrieve individual token representations across arbitrary delays; conversely, the LSTM maintained a coarser and more rapidly-decaying semantic gist of prior tokens, weighted toward the earliest items.

What Do We Know About Knowledge? Cover: AI@50--We Are Golden, by James Gary, New York, New York. Please contact AAAI for such permission. Electronic submissions should be submitted using the web-based submissions form. Submissions information is available at aimagazine.org.