LINKED: Eliciting, Filtering and Integrating Knowledge in Large Language Model for Commonsense Reasoning

TL;DR

A novel method named eliciting, filtering and integrating knowledge in large language model (LINKED) is proposed, which design a reward model to filter out the noisy knowledge and take the marginal consistent reasoning module to reduce invalid reasoning.

Abstract

Large language models (LLMs) sometimes demonstrate poor performance on knowledge-intensive tasks, commonsense reasoning is one of them. Researchers typically address these issues by retrieving related knowledge from knowledge graphs or employing self-enhancement methods to elicit knowledge in LLMs. However, noisy knowledge and invalid reasoning issues hamper their ability to answer questions accurately. To this end, we propose a novel method named eliciting, filtering and integrating knowledge in large language model (LINKED). In it, we design a reward model to filter out the noisy knowledge and take the marginal consistent reasoning module to reduce invalid reasoning. With our comprehensive experiments on two complex commonsense reasoning benchmarks, our method outperforms SOTA baselines (up to 9.0% improvement of accuracy). Besides, to measure the positive and negative impact of the injected knowledge, we propose a new metric called effectiveness-preservation score for the knowledge enhancement works. Finally, through extensive experiments, we conduct an in-depth analysis and find many meaningful conclusions about LLMs in commonsense reasoning tasks.

Figures (11)

• Figure 1: Some failed cases of traditional knowledge enhancement methods on complex commonsense reasoning tasks.
• Figure 2: The main architecture of our proposed method $\mathbb{LINKED}$.
• Figure 3: Comparison of different reasoning processes. The bars represent the probability distribution of options and the option marked in red indicates the final prediction in this sampling round.
• Figure 4: Comparison of the impact of different experimental factors on performance.
• Figure 5: Comparison of ES and PS on WinoGrande.