Self-conditioning pre-trained language models

We study the presence of expert units in pre-trained Transformer-based Language Models (TLMs), and how they can be used to condition text generation to contain specific concepts. We define expert units to be neurons that are able to detect a concept in the input with a given average precision. A concept is represented with a set of sentences that either do or do not contain the concept. Leveraging the OneSec dataset, we compile a dataset of 1344 concepts that allows diverse expert units in TLMs to be discovered. Our experiments demonstrate that off-the-shelf pre-trained TLMs can be conditioned on their own knowledge (self-conditioning) to generate text that contains a given concept. To this end, we intervene on the top expert units by fixing their output during inference, and we show experimentally that this is an effective method to condition TLMs. Our method does not require fine-tuning the model or using additional parameters, which allows conditioning large TLM with minimal compute resources. Furthermore, by intervening on a small number of experts in GPT2, we can achieve parity with respect to two concepts at generation time. The specific case of gender bias is explored, and we show that, for given contexts, gender parity is achieved while maintaining the model's perplexity.