Large language models (LLMs) demonstrate strong capabilities in understanding textual data and performing related tasks. However, a major challenge with LLMs is their high computational resource requirement, which makes them expensive to use. Small language models (SLMs) have been proposed as a more efficient alternative, showing promising results in handling textual data. Nevertheless, SLMs struggle when processing very large texts due to limitations in analyzing long inputs. In this paper, we propose using topic modeling (TM) to condense large texts into smaller, focused topic representations, which are then input to SLMs to extract overall topics from large documents. While TM produces topic distributions that often lack semantic interpretability, SLMs can assist by assigning meaningful labels to these topic clusters of words. Additionally, we investigate enriching the topic representations with knowledge graphs to provide more semantic context to the SLMs. We compare three proposed algorithms under consistent conditions to demonstrate that SLMs enhanced with structured information can achieve results comparable to LLMs. Experimental results show that the proposed approach balances computational efficiency and semantic accuracy, making it suitable for resource constrained environments. For instance, our method achieved a cosine similarity of 0.65 using an SLM performance previously reached only by models like ChatGPT-3.5 (OpenAI 2023). These findings suggest that combining topic modeling with lightweight language models is a promising direction for scalable semantic text analysis.