[Paper Review] Decoupling Knowledge from Memorization: Retrieval-augmented Prompt Learning

Xiang Chen, Lei Li+7 more|arXiv (Cornell University)|May 29, 2022

Topic Modeling26 citations

TL;DR

RetroPrompt builds an open-book knowledge-store from training data and uses retrieval to augment prompt learning, improving few-shot/zero-shot generalization and reducing reliance on memorization.

ABSTRACT

Prompt learning approaches have made waves in natural language processing by inducing better few-shot performance while they still follow a parametric-based learning paradigm; the oblivion and rote memorization problems in learning may encounter unstable generalization issues. Specifically, vanilla prompt learning may struggle to utilize atypical instances by rote during fully-supervised training or overfit shallow patterns with low-shot data. To alleviate such limitations, we develop RetroPrompt with the motivation of decoupling knowledge from memorization to help the model strike a balance between generalization and memorization. In contrast with vanilla prompt learning, RetroPrompt constructs an open-book knowledge-store from training instances and implements a retrieval mechanism during the process of input, training and inference, thus equipping the model with the ability to retrieve related contexts from the training corpus as cues for enhancement. Extensive experiments demonstrate that RetroPrompt can obtain better performance in both few-shot and zero-shot settings. Besides, we further illustrate that our proposed RetroPrompt can yield better generalization abilities with new datasets. Detailed analysis of memorization indeed reveals RetroPrompt can reduce the reliance of language models on memorization; thus, improving generalization for downstream tasks. Code is available in https://github.com/zjunlp/PromptKG/tree/main/research/RetroPrompt.

Motivation & Objective

• Motivate improving generalization in prompt learning beyond rote memorization.
• Decouple knowledge from memorization by creating an open-book knowledge-store from training data.
• Enable retrieval-driven augmentation during input, training, and inference to guide learning and prediction.

Proposed method

• Construct a dense retrieval-based knowledge-store from training instances as key-value pairs (h_hat_c, v) where h_hat_c is the [MASK] token embedding.
• Retrieve neural demonstrations by aggregating m nearest neighbors per class and injecting their representations into the input.
• Use k-nearest neighbors (kNN) to guide training by reweighting the cross-entropy loss based on the kNN-derived probability.
• Interpolate the kNN distribution with the PLM MLM output during inference to produce the final prediction.
• Refresh the knowledge-store asynchronously during training to keep embeddings aligned with model updates.
• Compare against strong prompt-learning baselines and conduct ablations to verify component contributions.

Experimental results

Research questions

• RQ1 Does retrieval-augmented prompt learning improve performance in few-shot and zero-shot settings compared to standard prompt tuning?
• RQ2 Can an internal, train-set-derived knowledge-store decouple knowledge from memorization and improve generalization to new domains/datasets?
• RQ3 How do neural demonstrations and kNN-guided training influence memorization and robustness to long-tail distributions?
• RQ4 What is the impact of updating/referencing the knowledge-store during training and inference on performance?

Key findings

• RetroPrompt achieves better performance than baselines across nine NLU tasks in few-shot and zero-shot settings.
• Retrieving from an open-book knowledge-store reduces reliance on memorization and improves generalization, including in fully supervised long-tail distributions.
• Neural demonstrations and kNN-guided training contribute notably to gains, especially in few-shot scenarios, with kNN-test offering limited gains on its own.
• RetroPrompt demonstrates stronger cross-domain generalization compared to baselines in cross-domain evaluations.
• Memorization analysis shows RetroPrompt reduces memorization scores relative to fine-tuning and traditional prompt-tuning.
• Ablation studies confirm that each component—neural demonstrations, kNN-guided training, and kNN-based prediction—contributes to performance gains.