Hindsight: Posterior-guided training of retrievers for improved open-ended generation

By Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei A. Zaharia, Christopher D. Manning. 2021

48 citations recorded by the source corpus.

Many text generation systems benefit from using a retriever to retrieve passages from a textual knowledge corpus (e.g., Wikipedia) which are then provided as additional context to the generator. For open-ended generation tasks (like generating informative utterances in conversations) many varied passages may be equally relevant and we find that existing methods that jointly train the retriever and generator underperform: the retriever may not find relevant passages even amongst the top-10 and hence the generator may not learn a preference to ground its generated output in them. We propose using an additional guide retriever that is allowed to use the target output and"in hindsight"retrieve relevant passages during training. We model the guide retriever after the posterior distribution Q of passages given the input and the target output and train it jointly with the standard retriever and the generator by maximizing the evidence lower bound (ELBo) in expectation over Q. For informative conversations from the Wizard of Wikipedia dataset, with posterior-guided training, the retriever finds passages with higher relevance in the top-10 (23% relative improvement), the generator's responses are more grounded in the retrieved passage (19% relative improvement) and the end-to-end system produces better overall output (6.4% relative improvement).

Subfields: Artificial Intelligence, Retrieval-Augmented Generation

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Retrieval-Augmented Generation

Citations in the corpus

Few-shot Learning with Retrieval Augmented Language Models (2022, methodology): “Paranjape et al. (2021) propose a method to train retrieval-augmented generators using a second “informed” retriever with access to the output, which the test-time retriever can be distilled from, and Hofstätter et al. (2022) recently proposed a training set ﬁltering/weighting approach to train…”
Re2G: Retrieve, Rerank, Generate (2022, background): “Since our submission to the KILT leaderboard for the Wizard of Wikipedia, a new system called Hindsight [Paranjape et al., 2021] achieved even better results on the generation metrics on that particular task.”
Multi-Task Retrieval-Augmented Text Generation with Relevance Sampling (2022, background): “Many of the recent papers improving RAG-style models optimized end-to-end processes (f.e. EMDR2 (Singh et al., 2021)), ensembling multiple modules (f.e. R2-D2 (Fajcik et al., 2021)), or creating multiple training loops to update the indexed documents multiple times (f.e. Hindsight (Paranjape et al.,”
Memory-Based Model Editing at Scale (2022, background): “Otherwise, the input is simply passed to the base model. model predictions do not always update to reﬂect the content of the prompts (Lewis et al., 2020; Paranjape et al., 2021).”
ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction (2021, background): “…years, dramatically advancing not only passage and document search (Nogueira and Cho, 2019) but also many knowledge-intensive NLP tasks like opendomain question answering (Guu et al., 2020), multi-hop claim verification (Khattab et al., 2021a), and open-ended generation (Paranjape et al., 2021).”
Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive NLP (2022, unknown): “We report the novel-F1 metric (nF1; Paranjape et al. 2022), which computes the F1 overlap between the system response and the ground truth while discounting common stopwords and terms present in the question (or earlier questions).”
Evidentiality-guided Generation for Knowledge-Intensive NLP Tasks (2021, background): “Retrievalaugmented generators achieve competitive performance across many different knowledge-intensive NLP tasks (Izacard and Grave, 2021b; Lewis et al., 2020b; Glass et al., 2021; Paranjape et al., 2021; Park et al., 2021).”