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.

Large language models exhibit high-level commonsense reasoning abilities, especially with enhancement methods like Chain-of-Thought (CoT). However, we find these CoT-like methods lead to a considerable number of originally correct answers turning wrong, which we define as the Toxic CoT problem. To interpret and mitigate this problem, we first utilize attribution tracing and causal tracing methods to probe the internal working mechanism of the LLM during CoT reasoning. Through comparisons, we prove that the model exhibits information loss from the question over the shallow attention layers when generating rationales or answers. Based on the probing findings, we design a novel method called RIDERS (Residual decodIng and sERial-position Swap), which compensates for the information deficit in the model from both decoding and serial-position perspectives. Through extensive experiments on multiple commonsense reasoning benchmarks, we validate that this method not only significantly eliminates Toxic CoT problems (decreased by 23.6%), but also effectively improves the model’s overall commonsense reasoning performance (increased by 5.5%).

Recent work shows large language models can be prompted to generate useful rationales for commonsense question answering (CQA), which can improve the performance of both themselves and other models. However, the cost of deployment and further tuning is relatively expensive for the large models. Some work explores to distill the the rationale-generation ability to convenient small-sized models, yet it typically requires human-authored QA instances during the distillation. In this paper, we propose a novel framework that leverages both knowledge graphs and large language models to synthesize rationale-augmented CQA data. Based on it, we train Leros, a model that can generate helpful rationales to assist generic QA models to accomplish unseen CQA tasks. Empirical results demonstrate Leros can substantially enhance the performance of QA models on five unseen CQA benchmarks, providing better gains than both same-sized counterpart models trained with downstream data and 10x larger language models. Our work reveals a novel way to integrate knowledge from both knowledge graphs and large language models into smaller models. The codes and synthesized resources are publicly available at https://github.com/wchrepo/leros.

Commonsense knowledge graphs (CKGs) are increasingly applied in various natural language processing tasks. However, most existing CKGs are limited to English, which hinders related research in non-English languages. Meanwhile, directly generating commonsense knowledge from pretrained language models has recently received attention, yet it has not been explored in non-English languages. In this paper, we propose a large-scale Chinese CKG generated from multilingual PLMs, named as **CN-AutoMIC**, aiming to fill the research gap of non-English CKGs. To improve the efficiency, we propose generate-by-category strategy to reduce invalid generation. To ensure the filtering quality, we develop cascaded filters to discard low-quality results. To further increase the diversity and density, we introduce a bootstrapping iteration process to reuse generated results. Finally, we conduct detailed analyses on CN-AutoMIC from different aspects. Empirical results show the proposed CKG has high quality and diversity, surpassing the direct translation version of similar English CKGs. We also find some interesting deficiency patterns and differences between relations, which reveal pending problems in commonsense knowledge generation. We share the resources and related models for further study.