Retrieval Augmentation Reduces Hallucination in Conversation

Abstract

Despite showing increasingly human-like conversational abilities, state-of-the-art dialogue models often suffer from factual incorrectness and hallucination of knowledge (Roller et al., 2021).In this work we explore the use of neural-retrieval-in-the-loop architectures -recently shown to be effective in open-domain QA (Lewis et al., 2020b; Izacard and Grave, 2021b) -for knowledge-grounded dialogue, a task that is arguably more challenging as it requires querying based on complex multi-turn dialogue context and generating conversationally coherent responses. We study various types of architectures with multiple components -retrievers, rankers, and encoder-decoders -with the goal of maximizing knowledgeability while retaining conversational ability. We demonstrate that our best models obtain state-of-the-art performance on two knowledge-grounded conversational tasks.The models exhibit open-domain conversational capabilities, generalize effectively to scenarios not within the training data, and, as verified by human evaluations, substantially reduce the well-known problem of knowledge hallucination in state-of-the-art chatbots.

Survey of Hallucination in Natural Language Generation

“…First, decoders could attend to the wrong part of the encoded input source [172]. This leads the generated output to contain mixed up facts between two similar entities [40,158]. Second, the design of decoding strategy itself can contribute to hallucinations.…”

“…For instance, Wang et al [184] proposes PARENT-T that simplifies PARENT by only using table content as the reference. Similarly, Knowledge F1 [158] -a variant of unigram F1 -is proposed for knowledge grounded dialogue tasks to measure the overlap between the model's generation and the knowledge used to ground the dialogue during dataset collection.…”

“…Due to the challenging and imperfect nature of the current automatic evaluation of hallucinations in NLG, human evaluation [154,158] is still one of the most commonly used approaches. There are two main forms of human evaluation: (1) scoring, where human annotators rate the hallucination level in a range; and (2) comparing, where human annotators compare the output texts with baselines or ground-truth references [165].…”

“…or on the other hand, hallucination [40,73,107,154,158], fact contradicting [129] are used in the human evaluation of hallucination to rate whether the generated text is in accord with the source input. Chen et al [22], Nie et al [130] use finer-grained metrics for intrinsic hallucination and extrinsic hallucination separately.…”

“…Consequently, a better semantic understanding helps alleviate the divergence issue from the source. For example, augmented with entity information [107], extracted relation triples from source document [20,73] obtained by Fact Description Extraction, synthetic data generated through replacement or perturbation [22,91], retrieved external knowledge [12,45,65,158,222], and retrieved similar training samples [13].…”