This paper presents Babelfy, a unified graph-based approach to Entity Linking (EL) and Word Sense Disambiguation (WSD). The approach combines the tasks by leveraging a semantic network, where candidate meanings are identified and then disambiguated using a densest subgraph heuristic to select the most coherent interpretations. Babelfy performs well on six datasets, including multilingual settings, achieving state-of-the-art results on both EL and WSD tasks. The method uses a semantic network, such as BabelNet, which integrates lexicographic and encyclopedic knowledge. It employs random walks with restart and triangle-based edge weighting to generate semantic signatures, which are then used to identify and disambiguate candidate meanings. The densest subgraph heuristic is used to select the most coherent interpretations, enabling the system to handle partial mentions and high ambiguity effectively. The approach is evaluated on various datasets, showing its effectiveness in both long and short texts, and its robustness across languages. The results demonstrate that the unified approach outperforms existing methods in both EL and WSD tasks.This paper presents Babelfy, a unified graph-based approach to Entity Linking (EL) and Word Sense Disambiguation (WSD). The approach combines the tasks by leveraging a semantic network, where candidate meanings are identified and then disambiguated using a densest subgraph heuristic to select the most coherent interpretations. Babelfy performs well on six datasets, including multilingual settings, achieving state-of-the-art results on both EL and WSD tasks. The method uses a semantic network, such as BabelNet, which integrates lexicographic and encyclopedic knowledge. It employs random walks with restart and triangle-based edge weighting to generate semantic signatures, which are then used to identify and disambiguate candidate meanings. The densest subgraph heuristic is used to select the most coherent interpretations, enabling the system to handle partial mentions and high ambiguity effectively. The approach is evaluated on various datasets, showing its effectiveness in both long and short texts, and its robustness across languages. The results demonstrate that the unified approach outperforms existing methods in both EL and WSD tasks.