A new arXiv paper proposes using pre-trained language models as zero-shot tools for two common social science tasks: record linkage and demographic inference from names. The author notes that most existing approaches depend on labeled training data that is difficult or costly to acquire, and argues that pre-trained models can bypass this requirement.
The paper focuses on name-based tasks, which are routine in social science research. By applying pre-trained language models directly, without task-specific fine-tuning, the method aims to provide a practical alternative for researchers who lack annotated datasets.
Because this is a single source, there are no conflicting findings to compare. The claims are presented as the author's demonstration rather than as an established benchmark.