Causal Reasoning of Stored Grain Pest Events Enhanced by Large Language Models

In the task of causal reasoning, some progress has been made in finding causal information between simple sentences in large-scale event graphs. However, when there is a complex causal relationship between multiple events in a sentence, the existing model is difficult to make full use of the effective information, resulting in poor results. To solve this problem, this paper proposes a Large Language Models-Event Graph Reasoner (LLMs-EGR). First, Large Language Models (LLMs) are used to excavate causal dependencies between events and form inference graphs. Secondly, the event graph reasoner is trained, and probabilistic logic is introduced to further quantify the causal strength relationship between events. Finally, the result events supported by the highest score inference graph are output as inference results. Comparative experiments were carried out in the field of stored grain pests event dataset SGPE and the causal reasoning baseline dataset (COPA, CosmosQA). The results show that LLMs-EGR can better excavate complex causality in text and effectively improve reasoning ability. Compared with the baseline model GrapeQA, the LLMs-EGR model improved the accuracy of complex inference problems by $\mathbf{2 . 6}$ percentage points.