Language models (LMs) can “memorize” information, i.e., encode training data in their weights in such a way that inference-time queries can lead to verbatim regurgitation of that data. This ability to extract training data can be problematic, for example, when data are private or sensitive. In this work, we investigate methods to mitigate memorization: three regularizer-based, three fine-tuning-based, and eleven machine unlearning-based methods, with five of the latter being new methods that we introduce. We also introduce TinyMem, a suite of small, computationally-efficient LMs for the rapid development and evaluation of memorization-mitigation methods. We demonstrate that the mitigation methods that we develop using TinyMem can successfully be applied to production-grade LMs, and we determine via experiment that: regularizer-based mitigation methods are slow and ineffective at curbing memorization; fine-tuning-based methods are effective at curbing memorization, but overly expensive, especially for retaining higher accuracies; and unlearning-based methods are faster and more effective, allowing for the precise localization and removal of memorized information from LM weights prior to inference. We show, in particular, that our proposed unlearning method BalancedSubnet outperforms other mitigation methods at removing memorized information while preserving performance on target tasks.

Despite significant progress in natural language processing, machine learning models require substantial expertannotated training data to perform well in tasks such as named entity recognition (NER) and entity relations extraction. Furthermore, NER is often more complicated when working with scientific text. For example, in polymer science, chemical structure may be encoded using nonstandard naming conventions, the same concept can be expressed using many different terms (synonymy), and authors may refer to polymers with ad-hoc labels. These challenges, which are not unique to polymer science, make it difficult to generate training data, as specialized skills are needed to label text correctly. We have previously designed polyNER, a semi-automated system for efficient identification of scientific entities in text. PolyNER applies word embedding models to generate entity-rich corpora for productive expert labeling, and then uses the resulting labeled data to bootstrap a context-based classifier. PolyNER facilitates a labeling process that is otherwise tedious and expensive. Here, we use active learning to efficiently obtain more annotations from experts and improve performance. Our approach requires just five hours of expert time to achieve discrimination capacity comparable to that of a state-of-the-art chemical NER toolkit.