Graph-Augmented Reasoning with Large Language Models for Tobacco Pest and Disease Management

This paper proposes a graph-augmented reasoning framework for tobacco pest and disease management that integrates structured domain knowledge into large language models. Building on GraphRAG, we construct a domain-specific knowledge graph and retrieve query-relevant subgraphs to provide relational evidence during answer generation. The framework adopts ChatGLM as the Transformer backbone with LoRA-based parameter-efficient fine-tuning, and employs a graph neural network to learn node representations that capture symptom-disease-treatment dependencies. By explicitly modeling diseases, symptoms, pesticides, and control measures as linked entities, the system supports evidence-aware retrieval beyond surface-level text similarity. Retrieved graph evidence is incorporated into the LLM input to guide generation toward domain-consistent recommendations and to mitigate hallucinated or inappropriate treatments. Experimental results show consistent improvements over text-only baselines, with the largest gains observed on multi-hop and comparative reasoning questions that require chaining multiple relations.

💡 Research Summary

The paper presents a graph‑augmented reasoning framework tailored for tobacco pest and disease management, integrating a domain‑specific knowledge graph with a large language model (LLM) to improve question answering (QA) performance, especially on multi‑hop and comparative queries. The authors begin by highlighting the challenges in tobacco cultivation in southern China: pest and disease outbreaks dramatically reduce yield, yet current decision‑making relies heavily on scattered expert knowledge and manual guidelines, making diagnosis slow and inconsistent. Knowledge graphs can organize such dispersed information into explicit entity‑relation structures, while LLMs such as ChatGLM excel at natural language understanding and generation. However, LLMs alone struggle with relational reasoning when the required dependencies are not explicitly modeled, leading to hallucinations or inappropriate treatment recommendations.

To address this, the authors build on GraphRAG, a recent approach that couples graph‑based retrieval with language generation. They construct a tobacco‑specific knowledge graph by aggregating data from agricultural literature, extension manuals, and expert‑curated resources. Nodes represent diseases, symptoms, pesticides, and control measures; edges encode typed relations such as has‑symptom, treated‑by, and prevention‑of. Each fact is stored as a triple (head, relation, tail). For embedding, they first train TransE, which treats each relation as a translation vector so that h + r ≈ t. This yields low‑dimensional vectors for entities and relations, learned via a margin‑based ranking loss with negative sampling. TransE provides a fast, similarity‑based retrieval foundation but does not capture higher‑order graph structure.

To incorporate neighborhood context, the authors refine the TransE embeddings with a two‑layer Graph Convolutional Network (GCN). The GCN propagates messages across adjacent nodes, normalizing by node degrees, and stacks multiple layers to aggregate multi‑hop information. Consequently, a disease node’s representation encodes not only its own attributes but also signals from linked symptoms and recommended treatments. These refined node vectors serve as query‑conditioned evidence in the subsequent retrieval step.

The language model backbone is ChatGLM, a transformer‑based LLM. For parameter efficiency, LoRA (Low‑Rank Adaptation) is applied, allowing fine‑tuning with a small set of additional low‑rank matrices. During inference, a user query is processed by the GraphRAG module, which identifies relevant entities in the graph and extracts their TransE + GCN embeddings. The query embedding and the graph embedding are concatenated into an augmented input vector (g_input) that is fed to the LLM alongside the natural‑language query. This design enables the self‑attention mechanism to align question tokens with graph‑derived evidence tokens, guiding generation toward answers that are grounded in explicit relational data rather than relying solely on the model’s parametric memory.

Experiments are conducted on a newly built tobacco pest‑disease QA dataset covering three question types: direct (single‑hop), multi‑hop, and comparative. Four configurations are compared: (1) ChatGLM alone, (2) KGE‑augmented ChatGLM (injecting only TransE embeddings), (3) text‑based Retrieval‑Augmented Generation (RAG), and (4) the proposed GraphRAG + ChatGLM. Evaluation metrics include Accuracy, Precision, Recall, and F1. Results (Table 1) show that GraphRAG + ChatGLM achieves the highest scores (Accuracy 90.1 %, Precision 92.3 %, Recall 88.2 %, F1 90.2 %). The gains are especially pronounced on multi‑hop and comparative questions, confirming that explicit relational evidence substantially improves chain‑of‑reasoning capabilities. Error analysis reveals that missing entities or relations in the graph and ambiguous user queries remain primary failure sources.

The discussion acknowledges limitations: constructing and maintaining a comprehensive knowledge graph is labor‑intensive; entity linking accuracy can bottleneck retrieval; and the current fusion strategy (simple vector concatenation) may not fully exploit the graph’s structural richness. Future work is outlined to expand graph coverage, adopt more expressive GNNs (e.g., GAT, RGCN), explore tighter integration mechanisms such as graph‑aware attention layers or path‑based prompting, and develop automated methods for graph enrichment.

In conclusion, the study demonstrates that augmenting LLMs with domain‑specific graph evidence yields a reliable, high‑performing decision‑support system for tobacco pest and disease management. By making symptom‑disease‑treatment relationships explicit and feeding query‑focused subgraphs into the generation process, the framework reduces hallucinations, improves multi‑step reasoning, and delivers recommendations that align with agricultural best practices. The approach offers a promising blueprint for other knowledge‑intensive agricultural domains where accurate, explainable AI assistance is critical.