# Graph-Augmented Reasoning with Large Language Models for Tobacco Pest and Disease Management
**Authors**: {lsy20, xyliu}@mails.cqjtu.edu.cn
## Abstract
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.
Graph-Augmented Reasoning with Large Language Models for Tobacco Pest and Disease Management
Siyu Li, Chenwei Song, Qi Zhou, Wan Zhou, Xinyi Liu School of Information Science and Engineering Chongqing Jiaotong University {lsy20, xyliu}@mails.cqjtu.edu.cn
## 1 Introduction
Tobacco is an important economic crop in southern China but is highly vulnerable to pests and diseases, which can substantially reduce yield and quality. Accurate diagnosis and timely selection of control measures are therefore critical for stable production and for avoiding unnecessary pesticide use. In practice, however, management decisions still rely heavily on expert experience and field observation, while relevant knowledge is scattered across manuals, extension guidelines, and research articles, making diagnosis and intervention slow, inconsistent, and difficult to scale.
Knowledge graphs provide an explicit entity–relation structure for organizing dispersed agricultural knowledge, while large language models (LLMs) offer strong language understanding and flexible reasoning (Paulheim, 2016; Brown et al., 2020; Wei et al., 2022). Their combination has become an effective paradigm for knowledge-intensive question answering and decision support (Lewis et al., 2020; Yasunaga et al., 2021). In tobacco pest–disease management, such integration is particularly appealing because key decisions require linking symptoms to diagnoses and diagnoses to appropriate treatments under practical constraints.
Despite these advantages, LLMs often struggle with domain reasoning when relational dependencies are not explicitly modeled (Sahoo et al., 2024). Prompt-based knowledge injection can be brittle, and text-only retrieval may surface relevant information without exposing the relations needed for multi-step reasoning. To address this limitation, we build on GraphRAG (Edge et al., 2024) and augment it with TransE embeddings (Bordes et al., 2013) and a GCN (Kipf, 2016) to better encode and exploit graph structure for downstream question answering.
Our approach emphasizes explicit relational modeling beyond entity mentions, using graph embeddings and GCN-based aggregation to provide structured, query-focused context to the LLM. As shown in Table ˜ 1, this design improves overall accuracy and yields the largest gains on multi-hop and comparative questions, where effective reasoning depends on chaining symptom–disease–treatment relations.
## 2 Related Work
#### Knowledge Graphs for Agricultural Pest and Disease Control.
Knowledge graphs have been widely used to organize agricultural pest–disease knowledge by linking entities (e.g., diseases, symptoms, treatments, and varieties) through explicit relations (Paulheim, 2016; Chithambarathanu and Jeyakumar, 2023). Crop-specific graphs (including tobacco and wheat) have supported decision support and management applications (Kamilaris and Prenafeta-Boldú, 2018). Recent work also studies temporal reasoning on such graphs to model evolving outbreaks and progression (Xiong et al., ; Xiong et al., 2024b).
#### Large Language Models and Graph-Enhanced Reasoning.
LLMs such as GPT-4 and LLaMA achieve strong generation and reasoning, but they still benefit from explicit structured knowledge when tasks require relational reasoning over entities (Brown et al., 2020; Wei et al., 2022; Yasunaga et al., 2021).
GraphRAG integrates graph-structured retrieval with language modeling by encoding entities and relations (e.g., via graph embeddings and GNNs) and injecting them into the generation process (Edge et al., 2024). Prior work shows that GCN/GAT-based node representations can guide LLMs to exploit relational paths and improve knowledge-intensive QA (Kipf, 2016; Yasunaga et al., 2021; Lewis et al., 2020). Related efforts further pursue tighter graph–LLM coupling for QA and commonsense reasoning (Yang et al., 2024; Xiong et al., 2024a, 2025).
#### Knowledge Graphs and LLMs in Agricultural Domains.
Combining knowledge graphs with LLMs for tobacco pest–disease applications is still underexplored. Prior work largely centers on domain knowledge representation or vision/ML-based diagnosis (Chithambarathanu and Jeyakumar, 2023), while recent studies investigate LLMs for broader agricultural knowledge tasks (Yin et al., 2022). We therefore study a GraphRAG-style integration to better leverage structured relations for tobacco-focused reasoning.
## 3 Methodology
We build a GraphRAG-style pipeline for tobacco pest–disease question answering that tightly couples structured graph evidence with neural generation. The method combines domain knowledge graph construction, TransE/GCN representation learning, and an LLM decoder to produce grounded responses. Figure-level details are omitted here; we describe the components and their interfaces in the following subsections.
### 3.1 Knowledge Graph Construction
We build a tobacco pest–disease knowledge graph by integrating information from agricultural literature, extension guidelines, and expert-curated resources. Nodes represent key entities (e.g., diseases, symptoms, pesticides, and field operations), while edges encode typed relations such as treatment-of, has-symptom, and prevention-of.
Formally, we model the graph as a directed graph $G=(V,E)$ , where $V$ is the entity set and $E$ is the relation set. Each fact is stored as a triple $(h,r,t)$ , which provides a unified interface for embedding, neighborhood aggregation, and query-time retrieval. For instance, (tobacco mosaic disease, treated by, spraying antiviral agents) links a disease to a control method; similar triples connect symptoms to diagnoses and diagnoses to recommended interventions.
### 3.2 Graph Embedding with TransE
We use TransE to embed entities and relations into a shared vector space, treating each relation as a translation operation. For a valid triple $(h,r,t)$ , the embeddings are encouraged to satisfy
$$
h+r≈t.
$$
This simple geometry provides an efficient way to encode large sets of symbolic facts and serves as the initialization for subsequent GCN refinement.
The training objective of TransE is defined as a margin-based ranking loss:
$$
L=∑_(h,r,t)∈ S∑_(h^\prime,r,t^{\prime)∈ S^\prime}\ell(h,r,t,h^\prime,t^\prime),
$$
$$
\ell=\Big[γ+d(h+r,t)-d(h^\prime+r,t^\prime)\Big]_+.
$$
where $S$ is the set of positive triples, $S^\prime$ is the set of negative triples generated by randomly replacing the head or tail entity, $γ$ is a margin hyperparameter, $d(·)$ denotes a distance function (e.g., L1 or L2 norm), and $[·]_+$ denotes the hinge loss. Negative sampling forces the model to separate true facts from plausible but incorrect alternatives. After training, entities and relations are represented as low-dimensional vectors that preserve relational semantics and support fast similarity-based retrieval.
### 3.3 Node Representation Learning with GCN
To incorporate neighborhood structure beyond pairwise triples, we refine TransE embeddings with a Graph Convolutional Network (GCN). This step propagates information over the local subgraph, allowing each entity to encode both its own attributes and relational context.
Let $h_i^(l)$ denote the representation of node $i$ at the $l$ -th GCN layer. The update rule of GCN is given by:
$$
h_i^(l+1)=σ≤ft(∑_j∈N(i)\frac{1}{√{\deg(i)\deg(j)}}W^(l)h_j^(l)\right),
$$
where $N(i)$ denotes the set of neighboring nodes of $i$ , $\deg(i)$ is the degree of node $i$ , $W^(l)$ is a trainable weight matrix, and $σ(·)$ is an activation function such as ReLU. The normalization term $\frac{1}{√{\deg(i)\deg(j)}}$ is used to stabilize training and prevent numerical explosion. In practice, this operator can be viewed as a normalized message-passing step over adjacent entities.
By stacking multiple GCN layers, the model aggregates multi-hop context so disease nodes can absorb signals from linked symptoms and control measures. These refined representations are later used as query-conditioned evidence in the GraphRAG fusion stage.
### 3.4 Integration with Large Language Models
We use a Transformer-based LLM (e.g., ChatGLM/LLaMA) as the generative backbone. For an input sequence
$$
X=(x_1,x_2,…,x_n),
$$
the model factorizes the likelihood in an autoregressive manner:
$$
P(X)=∏_i=1^np(x_i\mid x_<i).
$$
In our setting, the input includes both the user query and retrieved evidence so the decoder can condition on external knowledge during generation.
Transformer inference is based on self-attention:
$$
Attention(Q,K,V)=softmax≤ft(\frac{QK^⊤}{√{d_k}}\right)V,
$$
with standard query/key/value projections and positional encoding. This mechanism allows the model to align question tokens with graph-derived evidence tokens when producing the final answer.
### 3.5 GraphRAG-Based Fusion
Given a user query, GraphRAG identifies relevant entities and retrieves a small, query-focused subgraph as evidence. For “How to prevent tobacco mosaic disease?”, we locate the disease entity, obtain its TransE embedding $e_d$ , and compute a context-aware representation $g_d$ via GCN aggregation.
The query embedding and graph embedding are concatenated to form an augmented input:
$$
g_input=[e_d;g_d],
$$
which is then provided to the LLM as additional context. This fusion encourages the model to base its response on retrieved relations (e.g., recommended antiviral agents and sanitation measures) rather than relying solely on parametric memory.
### 3.6 Summary
Overall, the proposed methodology combines (i) symbolic relational knowledge from a domain graph and (ii) neural generation from an LLM. Graph-augmented retrieval supplies targeted, structured evidence, while TransE and GCN provide compact representations for efficient matching and reasoning. Together, these components improve reliability for tobacco pest–disease decision support.
## 4 Experiments
We evaluate the proposed approach on tobacco pest–disease question answering to quantify the benefit of structured graph evidence for domain reasoning (Lewis et al., 2020; Yasunaga et al., 2021). In addition to overall performance, we emphasize reasoning-heavy subsets (multi-hop and comparison) where relational structure is most critical.
Table 1: Comparison of Different Methods on Knowledge Reasoning Tasks for Tobacco Pest and Disease Control
| ChatGLM KGE + ChatGLM RAG + ChatGLM | 75.2 82.5 85.7 | 78.5 85.3 87.9 | 72.1 79.8 83.2 | 75.2 82.4 85.5 |
| --- | --- | --- | --- | --- |
| GraphRAG + ChatGLM | 90.1 | 92.3 | 88.2 | 90.2 |
### 4.1 Experimental Setup
We construct a tobacco pest–disease knowledge graph from expert resources and literature, and build a QA dataset covering direct, multi-hop, and comparative questions (Kamilaris and Prenafeta-Boldú, 2018; Chithambarathanu and Jeyakumar, 2023; Yasunaga et al., 2021; Wei et al., 2022). We split the data into training/dev/test sets and tune hyperparameters on the development set. For graph representation learning, we use TransE to initialize entity/relation embeddings and refine node representations with a two-layer GCN (Bordes et al., 2013; Kipf, 2016). For the language model, we adopt ChatGLM and apply LoRA for parameter-efficient tuning, then fuse retrieved graph evidence with the LLM input under GraphRAG (Hu et al., 2022; Edge et al., 2024; Reimers and Gurevych, 2019).
### 4.2 Baselines and Evaluation Metrics
We compare against (i) ChatGLM without external knowledge, (ii) KGE-assisted variants that inject entity embeddings, and (iii) standard text-based RAG (Lewis et al., 2020; Sahoo et al., 2024). Performance is reported with Accuracy, Precision, Recall, and F1-score (Chithambarathanu and Jeyakumar, 2023), computed on the same test set for all methods.
### 4.3 Results and Analysis
As shown in Table ˜ 1, GraphRAG+ChatGLM achieves the best results across all metrics. Compared with ChatGLM alone, adding graph retrieval and GCN refinement yields a large improvement, suggesting that explicit relational evidence reduces reliance on fragile parametric recall. Improvements are most pronounced on multi-hop and comparative questions, where correct answers require chaining symptom–disease–treatment relations. In these cases, text-only RAG can retrieve superficially similar passages but still miss key relational constraints, while graph-based retrieval offers more targeted evidence.
### 4.4 Discussion
Overall, graph-augmented retrieval and GCN-based representation learning provide a practical boost for tobacco pest–disease reasoning. Typical remaining errors come from missing entities/relations in the graph or ambiguous user questions, indicating that coverage and entity linking are important bottlenecks. In future work, we plan to expand the graph and dataset scale, incorporate stronger GNN encoders, and explore more principled fusion strategies for injecting graph evidence into the LLM.
## 5 Conclusion
We presented a graph-augmented LLM framework for tobacco pest–disease question answering that integrates GraphRAG-style retrieval with GCN-enhanced graph representations. By explicitly modeling symptom–disease–treatment relations and incorporating query-relevant subgraphs into the generation process, the framework provides structured relational evidence beyond surface-level text retrieval. Experimental results show consistent improvements over LLM-only and text-based RAG baselines, with particularly strong gains on multi-hop and comparative reasoning tasks that require chaining information across multiple entities. These findings suggest that making relational structure explicit can effectively support more reliable and domain-consistent reasoning in agricultural decision support systems.
## References
- A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, and O. Yakhnenko (2013) Translating embeddings for modeling multi-relational data. Advances in neural information processing systems 26. Cited by: §1, §4.1.
- T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. (2020) Language models are few-shot learners. Advances in neural information processing systems 33, pp. 1877–1901. Cited by: §1, §2.
- M. Chithambarathanu and M. Jeyakumar (2023) Survey on crop pest detection using deep learning and machine learning approaches. Multimedia Tools and Applications 82 (27), pp. 42277–42310. Cited by: §2, §2, §4.1, §4.2.
- D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, D. Metropolitansky, R. O. Ness, and J. Larson (2024) From local to global: a graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130. Cited by: §1, §2, §4.1.
- E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al. (2022) Lora: low-rank adaptation of large language models.. ICLR 1 (2), pp. 3. Cited by: §4.1.
- A. Kamilaris and F. X. Prenafeta-Boldú (2018) Deep learning in agriculture: a survey. Computers and electronics in agriculture 147, pp. 70–90. Cited by: §2, §4.1.
- T. Kipf (2016) Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907. Cited by: §1, §2, §4.1.
- P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, et al. (2020) Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, pp. 9459–9474. Cited by: §1, §2, §4.2, §4.
- H. Paulheim (2016) Knowledge graph refinement: a survey of approaches and evaluation methods. Semantic web 8 (3), pp. 489–508. Cited by: §1, §2.
- N. Reimers and I. Gurevych (2019) Sentence-bert: sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084. Cited by: §4.1.
- P. Sahoo, A. K. Singh, S. Saha, V. Jain, S. Mondal, and A. Chadha (2024) A systematic survey of prompt engineering in large language models: techniques and applications. arXiv preprint arXiv:2402.07927. Cited by: §1, §4.2.
- J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al. (2022) Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, pp. 24824–24837. Cited by: §1, §2, §4.1.
- S. Xiong, A. Payani, R. Kompella, and F. Fekri (2024a) Large language models can learn temporal reasoning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 10452–10470. Cited by: §2.
- S. Xiong, A. Payani, Y. Yang, and F. Fekri (2025) Deliberate reasoning in language models as structure-aware planning with an accurate world model. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 31900–31931. Cited by: §2.
- [15] S. Xiong, Y. Yang, F. Fekri, and J. C. Kerce TILP: differentiable learning of temporal logical rules on knowledge graphs. In The Eleventh International Conference on Learning Representations, Cited by: §2.
- S. Xiong, Y. Yang, A. Payani, J. C. Kerce, and F. Fekri (2024b) Teilp: time prediction over knowledge graphs via logical reasoning. In Proceedings of the AAAI conference on artificial intelligence, Vol. 38, pp. 16112–16119. Cited by: §2.
- Y. Yang, S. Xiong, A. Payani, E. Shareghi, and F. Fekri (2024) Harnessing the power of large language models for natural language to first-order logic translation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 6942–6959. Cited by: §2.
- M. Yasunaga, H. Ren, A. Bosselut, P. Liang, and J. Leskovec (2021) QA-gnn: reasoning with language models and knowledge graphs for question answering. arXiv preprint arXiv:2104.06378. Cited by: §1, §2, §2, §4.1, §4.
- D. Yin, L. Dong, H. Cheng, X. Liu, K. Chang, F. Wei, and J. Gao (2022) A survey of knowledge-intensive nlp with pre-trained language models. arXiv preprint arXiv:2202.08772. Cited by: §2.