Zero-shot controllable text generation

Project description

From the original project page. Results and time references describe that version of the work.

We explore the use of contrasting expert language models for the use of topical control on language generation. As a result, we examine methods to extract linguistically significant meaning from self-organizing structures trained on a large corpus of Wikipedia-drawn text summaries and their corresponding categories. Our findings show extracting such meaning is difficult and needs further work, with the contrastive experts expressing opposing probability adjustments that simply cancel each other out. Further, when using only our positive expert as opposed to the contrastive experts, we observe metric improvements for our in-domain data but a failure to match pre-prompting baselines on out-of-domain data. Our work is a novel exploration of self-organizing structures for this purpose and presents a starting point for such methods going forward.