Efficient Topic Extraction via Graph-Based Labeling: A Lightweight Alternative to Deep Models
To address the high computational cost and poor interpretability of deep models in topic modeling, this paper proposes a lightweight graph-enhanced topic labeling method. Unlike LLM-dependent approaches, it constructs a semantic word graph over pre-trained word embeddings, leverages graph propagation to expand topic vocabularies and uncover structured semantic relationships among words, and generates concise, semantically coherent topic labels. Experiments on two standard benchmarks demonstrate that our method significantly outperforms traditional baselines (e.g., LDA+KeyBERT) in label quality—measured by BERTScore and cosine similarity—and matches ChatGPT-3.5’s performance, while achieving 3–5× faster inference and reducing memory consumption by ~70%. Our core contribution is the first integration of graph-structured modeling into topic labeling, enabling efficient, low-resource, and highly interpretable automatic label generation.