Retrieval-augmented generation models offer many benefits over standalone language models: besides a textual answer to a given query they provide provenance items retrieved from an updateable knowledge base. However, they are also more complex systems and need to handle long inputs. In this work, we introduce FiDLight to strongly increase the efficiency of the state-of-the-art retrieval-augmented FiD model, while maintaining the same level of effectiveness. Our FiD-Light model constrains the information flow from the encoder (which encodes passages separately) to the decoder (using concatenated encoded representations). Furthermore, we adapt FiD-Light with re-ranking capabilities through textual source pointers, to improve the top-ranked provenance precision. Our experiments ona diverse set of seven knowledge intensive tasks (KILT) show FiD-Light consistently improves the Pareto frontier between query latency and effectiveness. FiD-Light with source pointing sets substantial new state-of-the-art results on six KILT tasks for combined text generation and provenance retrieval evaluation, while maintaining reasonable efficiency (DeepL)Search augmented generative models offer many advantages over stand-alone language models. In addition to textual answers to a given query, it provides provenance items taken from an updatable knowledge base. However, it is a more complex system and needs to handle longer inputs. By introducing FiDLight, we significantly improve the efficiency of a state-of-the-art search augmented FiD model while maintaining the same level of effectiveness. Our FiD-Light model constrains the flow of information from the encoder (which encodes the passages separately) to the decoder (which concatenates the encoded representations). In addition, we adapted FiD-Light to re-rank with text source pointers to improve top-ranked provenance accuracy; experiments on a diverse set of seven knowledge-intensive tasks (KILTs) show that FiD-Light consistently improves the query latency vs. the effectiveness of the Pareto frontier was shown to be consistently improved. FiD-Light with source pointing showed substantially new state-of-the-art results in six KILT tasks for evaluation combining text generation and provenance retrieval, while maintaining reasonable efficiency.
2209.14290 FiD-Light: Efficient and Effective Retrieval-Augmented Text Generation
This page is auto-translated from [/nishio/FID-LIGHT: EFFICIENT AND EFFECTIVE RETRIEVAL-AUGMENTED TEXT GENERATION](https://scrapbox.io/nishio/FID-LIGHT: EFFICIENT AND EFFECTIVE RETRIEVAL-AUGMENTED TEXT GENERATION) using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.