Guiding ChatGPT to Generate Salient Space Summaries
Authors: Jun Gao, Ziqiang Cao, Shaoyao Huang, Luozheng Qin, Chunhui Ai
Abstract: ChatGPT is instruct-tuned to generate widespread and human-expected content material materials to align with human need by Reinforcement Learning from Human Recommendations (RLHF), within the meantime resulting in generated responses not salient adequate. Resulting from this reality, on this case, ChatGPT may fail to meet space requirements in zero-shot settings, leading to poor ROUGE scores. Impressed by the In-Context Learning (ICL) and retelling potential of ChatGPT, this paper proposes PADS, a textbf{P}ipeline for textbf{A}ssisting ChatGPT in textbf{D}omain textbf{S}ummarization. PADS consists of a retriever to retrieve comparable examples from corpora and a rank model to rerank the various candidate summaries generated by ChatGPT. Notably, given an inference doc, we first retrieve an in-context demonstration by means of the retriever. Then, we require ChatGPT to generate okay candidate summaries for the inference doc at a time beneath the steering of the retrieved demonstration. Lastly, the rank model independently scores the okay candidate summaries in accordance with their top quality and selects the optimum one. We extensively uncover dense and sparse retrieval methods to pick environment friendly demonstrations for reference and successfully follow the rank model to copy the usual of candidate summaries for each given summarized doc. Furthermore, PADS incorporates merely 400M trainable parameters originating from the rank model and we merely collect 2.5k info to teach it. We take into account PADS on 5 datasets from utterly totally different domains, and the consequence signifies that each module in PADS is devoted to efficiently guiding ChatGPT to generate salient summaries changing into utterly totally different space requirements. Notably, inside the well-liked summarization dataset Gigaword, PADS achieves over +8 obtain on ROUGE-L, in distinction with the naive ChatGPT inside the zero-shot setting. footnote{Our code might be discovered at url{https://github.com/jungao1106/PADS}}
