The scale and speed of online disinformation increasingly exceed the capacity of manual verification and fact-checking. This challenge is especially acute in active conflict settings, where disinformation can shape public opinion, undermine institutional trust, and influence humanitarian outcomes. This paper investigates whether semantic similarity links can improve graph-based detection of pro-Russian disinformation narratives in the Ukrainian-Russian Telegram ecosystem.

We construct a heterogeneous directed graph of 81,369 Telegram messages from 95 public channels, representing channel-message provenance, explicit forwarding relations, and semantic similarity edges derived from multilingual BAAI/bge-m3 embeddings. Message-level narrative labels are obtained through weak supervision using few-shot LLM prompting against the TeleNarratives dataset taxonomy. We train a two-layer HeteroGraphSAGE model with LSTM aggregators under four edge configurations and compare it with a RoBERTa text-only baseline using a channel-aware chronological split.

Results show that semantic similarity edges provide an additional signal for narrative detection beyond explicit forwarding structure alone. The best full-graph configuration, using channel-posting and semantic similarity edges, achieves F1 = 0.82, MCC = 0.60, and PR-AUC = 0.81, closely matching the RoBERTa baseline. A restricted subgraph containing only messages involved in semantic similarity relations reaches F1 = 0.89 and PR-AUC = 0.94, suggesting that semantic-neighborhood structure is highly informative for identifying narrative-bearing messages. These results position semantic graph construction as a useful complement to forwarding-based diffusion analysis for monitoring disinformation narratives on Telegram.