Low-resource languages present a persistent challenge in machine translation (MT) due to the scarcity of digital resources, such as linguistic tools and parallel corpora. Consequently, research contributions addressing these gaps often go unnoticed, hindering progress in the field. To address this, we propose a novel approach that leverages graph networks, where embedding of a research paper serve as nodes, connected by edges based on cosine similarity (threshold ≥ 0.7). To optimize the graph structure, we compare two embedding models: Sentence-BERT (S-BERT) and Word2Vec. Results indicate a significant difference between the models (p-value: \(3.658 \times 10^{ - 5}\) , lch’s t-test), with S-BERT generating more stable embeddings. Furthermore, integrating the graph network into Gemini language model enhances inference generation, producing more specific and detailed insights. This framework effectively identifies key research trends, advancing low-resource MT analysis.