Controllable text generation: toward non-toxic, unbiased and factual language models for sensitive applications

we will investigate the following methodologies as a starting point:
1. Parameter-e_cient paradigms - which allow fine-tuning a pretrained model for a
downstream task without modifying all of its parameters - are not only more computationally efficient: they have also been found to outperform full model tuning in many cases [12, 13, 14]. These paradigms include prompt tuning [12], adapters [13] and prefix tuning [14]. It is worth noting that recent papers have investigated all three approaches in relation to the detoxification and debiasing of language models, with encouraging results [15, 16]. We aim to investigate these paradigms and build upon the existing research.

2. Retrieval-augmented generation has been shown to achieve state-of-the-art-performance
in a number of benchmarks [17]. This approach augments the knowledge implicitly stored in a model's parameters with a knowledge retriever that attends to an external corpus, such as Wikipedia or a knowledge graph [18, 19]. Since these corpora can encode facts with historical and scientific accuracy as well as societal norms and values [20, 21, 22], it is worth investigating whether retrieval-augmented approaches can result in less biased and less toxic models.