How Much Knowledge Can You Pack Into the Parameters of a Language Model? (2002.08910v4)

Abstract: It has recently been observed that neural LLMs trained on unstructured text can implicitly store and retrieve knowledge using natural language queries. In this short paper, we measure the practical utility of this approach by fine-tuning pre-trained models to answer questions without access to any external context or knowledge. We show that this approach scales with model size and performs competitively with open-domain systems that explicitly retrieve answers from an external knowledge source when answering questions. To facilitate reproducibility and future work, we release our code and trained models at https://goo.gle/t5-cbqa.

• The paper demonstrates that neural language models can answer open-domain questions by relying solely on pre-trained knowledge.
• It employs various T5 model sizes and salient span masking pre-training to reveal a strong correlation between model scale and QA accuracy.
• Results indicate that closed-book QA systems offer practical efficiency and inspire further research into scalable and interpretable pre-training methods.

How Much Knowledge Can You Pack Into the Parameters of a LLM?

In their paper, "How Much Knowledge Can You Pack Into the Parameters of a LLM?", Roberts et al. present an investigation into the capacity of neural LLMs to store and retrieve knowledge embedded within their parameters. By leveraging the Text-to-Text Transfer Transformer (T5) models, the authors offer a novel perspective on question answering (QA) tasks by demonstrating the potential for these models to operate without external knowledge bases, an approach they term "closed-book question answering".

Core Contributions

The paper addresses two primary research questions:

1. Can neural LLMs answer open-domain questions without external context?
2. How does model size impact the performance in storing and retrieving knowledge?

To investigate these questions, the authors fine-tune pre-trained T5 models on various QA datasets: Natural Questions, WebQuestions, and TriviaQA. Importantly, rather than providing supplementary documents containing the answers, the model is coerced to rely solely on the information assimilated during pre-training.

Experimental Setup

Model Variants

The authors evaluate multiple T5 model variants differing in parameter count:

• T5-Base (220M parameters)
• T5-Large (770M parameters)
• T5-3B (3B parameters)
• T5-11B (11B parameters)

Additionally, the paper introduces "T5.1.1" checkpoints which were pre-trained exclusively on unlabeled data for enhanced robustness.

Datasets and Evaluation

The selected datasets (Natural Questions, WebQuestions, TriviaQA) each contain questions typically used in open-domain QA, but for the purpose of their experiments, only the questions themselves are used, divorced from any additional contextual information. Evaluation metrics include exact match and recall scores.

Salient Span Masking

To optimize the models, the authors leverage the "salient span masking" (SSM) pre-training objective from Guu et al., which focuses on reconstructing masked-out spans from sentences rich in entities or dates. This approach was observed to produce significant gains in QA performance relative to standard span corruption.

Results

The experiments demonstrate several key findings:

1. Model Size and Performance: There is a clear, positive correlation between model size and performance across all datasets, with the largest models (T5-11B and T5.1.1-XXL) achieving the highest accuracy.
2. SSM Pre-Training Effectiveness: Employing the SSM objective yielded notable performance boosts, particularly with large model variants. For example, T5.1.1-XXL with SSM achieved state-of-the-art results on WebQuestions, outperforming prior approaches which utilized external knowledge retrieval.
3. Error Analysis: Manual inspection of predicted answers revealed a substantial incidence of false negatives attributed to phrasing mismatches, incomplete annotations, and unanswerable questions. Correctly accounting for these factors would further elevate performance scores.

Discussions and Implications

The results emphasize the efficiency of LLMs in internalizing vast amounts of knowledge. This approach challenges traditional dependency on external knowledge retrieval systems, suggesting that adequately pre-trained models can serve as robust, self-contained QA systems.

Practical Implications

One significant implication is the resource efficiency in deploying closed-book QA systems. By eliminating the need for extensive external knowledge retrieval and processing, costs related to computation and memory can be substantially mitigated. This is especially pertinent in scenarios where rapid real-time responses are critical.

Theoretical Implications

From a theoretical standpoint, the work underscores the importance of pre-training objectives. The superiority of SSM highlights that targeted pre-training can significantly influence the model's ability to store and retrieve knowledge effectively.

Future Directions

The authors outline several avenues for future research:

1. Efficient Model Architectures: Developing models that can retain similar or better performance with reduced computational overhead.
2. Interpretability: Investigating methods to elucidate how models store and retrieve specific knowledge, enhancing interpretability and trustworthiness.
3. Knowledge Updates: Examining mechanisms to dynamically update or remove knowledge within models post pre-training.
4. Broader QA Tasks: Expanding evaluation to QA tasks necessitating complex reasoning and multi-hop inference, such as the DROP dataset.

Conclusion

Roberts et al.'s work presents compelling evidence that LLMs are capable of storing and retrieving substantial amounts of knowledge within their parameters. By sidestepping the need for external context, these models demonstrate a paradigm shift in QA systems, fostering both practical efficiency and theoretical enrichment in Natural Language Processing research.