[Paper Review] Can Generative Pre-trained Language Models Serve as Knowledge Bases for Closed-book QA?

Cunxiang Wang, Pai Liu+1 more|arXiv (Cornell University)|Jun 3, 2021

Topic Modeling21 references4 citations

TL;DR

This paper investigates whether generative pre-trained language models like BART can serve as knowledge bases for closed-book question answering by constructing a new SQuAD-based benchmark with traceable knowledge sources. Despite high factual memorization in controlled settings, BART struggles with closed-book QA due to poor knowledge retention and retrieval, though decoupling pre-training from fine-tuning and prompting for knowledge recall show promise in improving performance.

ABSTRACT

Recent work has investigated the interesting question using pre-trained language models (PLMs) as knowledge bases for answering open questions. However, existing work is limited in using small benchmarks with high test-train overlaps. We construct a new dataset of closed-book QA using SQuAD, and investigate the performance of BART. Experiments show that it is challenging for BART to remember training facts in high precision, and also challenging to answer closed-book questions even if relevant knowledge is retained. Some promising directions are found, including decoupling the knowledge memorizing process and the QA finetune process, forcing the model to recall relevant knowledge when question answering.

Motivation & Objective

• To evaluate whether generative pre-trained language models like BART can function as internal knowledge bases for closed-book QA without external retrieval.
• To address the limitations of prior work, which relied on small benchmarks with high test-train overlap, by constructing a new benchmark using SQuAD with traceable Wikipedia passages.
• To isolate and analyze the two key challenges in closed-book QA: (1) memorizing factual knowledge accurately, and (2) retrieving and using that knowledge to answer questions.
• To investigate whether fine-tuning for QA negatively affects factual knowledge retention and whether alternative training strategies can improve performance.

Proposed method

• Constructed a new closed-book QA benchmark using SQuAD, where each question is paired with its source Wikipedia passage to enable traceable knowledge evaluation.
• Performed language model fine-tuning (LM-finetuning) on BART using related passages to enhance factual knowledge memorization.
• Evaluated knowledge retention via a 'reciting' task—measuring the model's ability to reconstruct factual spans from passages—under varying numbers of training passages.
• Conducted QA fine-tuning on the LM-fine-tuned BART models to assess downstream closed-book QA performance, even when training and test questions are grounded in the same knowledge.
• Explored data augmentation techniques, such as injecting related passages into test-time outputs, to improve knowledge retrieval during inference.
• Proposed and evaluated a decoupled training strategy, separating LM-finetuning and QA-finetuning to preserve factual knowledge during task-specific adaptation.

Experimental results

Research questions

• RQ1 Can generative pre-trained language models like BART effectively memorize factual knowledge from a diverse set of passages when fine-tuned in a language modeling objective?
• RQ2 Does the performance of BART on closed-book QA degrade despite high factual knowledge retention, indicating a failure in knowledge retrieval or utilization?
• RQ3 To what extent does QA fine-tuning interfere with the factual knowledge memorized during language model fine-tuning?
• RQ4 Can simple data augmentation, such as injecting relevant passages into test inputs, improve closed-book QA performance by aiding knowledge retrieval?
• RQ5 Does decoupling the language modeling fine-tuning and QA fine-tuning processes lead to better retention of factual knowledge and improved QA accuracy?

Key findings

• BART achieves only 1.5% accuracy on the SQuAD-based closed-book QA benchmark, despite high performance on other datasets with high test-train overlap.
• When fine-tuned on 500 passages, BART memorizes 66% of the required factual knowledge; this drops to just 4% when fine-tuned on 5,000 passages, indicating severe catastrophic forgetting.
• Even when BART retains most factual knowledge through LM-finetuning, it still fails to answer the majority of closed-book questions after QA fine-tuning, showing a fundamental difficulty in knowledge utilization.
• QA fine-tuning negatively impacts the model’s ability to recall facts in the reciting task, suggesting a trade-off between task-specific adaptation and factual retention.
• Injecting relevant passages into the test input improves BART’s performance by enabling better knowledge retrieval, indicating that external prompt engineering can compensate for internal knowledge gaps.
• Decoupling LM-finetuning and QA-finetuning helps preserve factual knowledge during task-specific adaptation, offering a promising path to improve closed-book QA performance.