P-Adapters: Robustly Extracting Factual Information from Language Models with Diverse Prompts (2110.07280v2)

Abstract: Recent work (e.g. LAMA (Petroni et al., 2019)) has found that the quality of the factual information extracted from LLMs depends on the prompts used to query them. This inconsistency is problematic because different users will query LLMs for the same information using different wording, but should receive the same, accurate responses regardless. In this work we aim to address this shortcoming by introducing P-Adapters: lightweight models that sit between the embedding layer and first attention layer of LLMs. They take LLM embeddings as input and output continuous prompts that are used to query the LLM. Additionally, we investigate Mixture of Experts (MoE) models that learn a set of continuous prompts ("experts") and select one to query the LLM. They require a separate classifier trained on human-annotated data to map natural language prompts to the continuous ones. P-Adapters perform comparably to the more complex MoE models in extracting factual information from BERT and RoBERTa while eliminating the need for additional annotations. P-Adapters show between 12-26% absolute improvement in precision and 36-50% absolute improvement in consistency over a baseline of only using natural language queries. Finally, we investigate what makes P-Adapters successful and conclude that a significant factor is access to the LLM's embeddings of the original natural language prompt, particularly the subject of the entity pair being queried.

• The paper presents P-Adapters, lightweight modules that enhance factual extraction from LLMs by transforming embeddings into continuous prompts.
• Their methodology inserts P-Adapters between the embedding and attention layers, eliminating the need for extra annotations required by MoE models.
• Experimental results show 12-26% precision and 36-50% consistency gains, demonstrating robust performance with diverse natural language prompts.

Overview of P-Adapters: Robust Information Extraction from LLMs

The paper "P-Adapter{s}: Robustly Extracting Factual Information from LLMs with Diverse Prompts" addresses the challenge of inconsistent factual information retrieval from LLMs due to variability in user prompts. This is critical, as varying prompts for the same information should yield the consistent, accurate results. The authors propose P-Adapter{s}, lightweight models designed to improve this consistency and accuracy.

Methodology

P-Adapter{s} are situated between the embedding layer and the first attention layer of an LLM. They transform LLM embeddings into continuous prompts, optimizing the query process. The research contrasts these adapters with the Mixture of Experts (MoE) models, which require a separately trained classifier to map natural language prompts to continuous prompts. The P-Adapter{s} achieve comparable performance to MoE models without requiring additional relation annotations, thereby simplifying the querying process.

The paper evaluates P-Adapters against three configurations with BERT and RoBERTa models:

1. In-Domain (ID) Testing: Evaluates generalization within the template and objects.
2. Out-of-Domain (OOD) Prompts: Tests new natural language prompts.
3. OOD Objects: Tests against a different entity distribution than training data.
4. OOD Keyboard Errors: Assesses robustness to typographic errors.

Key Results

The P-Adapter{s} show significant performance improvements:

• Precision Improvement: They achieve 12-26% absolute improvements in precision over natural language queries alone.
• Consistency Gain: They realize 36-50% improvements in consistency, which indicates reliable predictions across varied natural language prompts.

Specifically, the most challenging evaluation was found to be the OOD Objects, highlighting the potential for overfitting to the distribution of objects.

Insights and Implications

The P-Adapters provide insights into the crucial role of maintaining access to the LLM's original embeddings, particularly those of the subject entity. This contrasts with the assumption that subjects during extraction might matter less. The research demonstrates that leveraging the unmodified embeddings of subject terms notably boosts performance.

From a practical standpoint, P-Adapter{s} offer a methodology for reducing the dependency on discrete prompt engineering and heavy annotation, presenting a low-parameter alternative in fact extraction. They support a user-friendly interface, making LLMs more efficient as knowledge bases without the overhead of extensive tuning.

Future Directions

The development of P-Adapters opens several avenues for future research. Further exploration could focus on enhancing adaptability and robustness to a broader range of unstructured prompts and error conditions. Additionally, deploying P-Adapters in real-world applications could contribute to examining and mitigating any biases learned during LLM pretraining sessions.

In conclusion, while P-Adapter{s} are not without their limitations, particularly concerning precision in out-of-distribution tasks, they represent a promising step towards more consistent and accurate information retrieval from LLMs. As AI systems increasingly serve as information sources, such models become pivotal in ensuring reliability and user satisfaction.