<p>A vast amount of vessel trajectory (VT) data is generated during vessel movements. This data contains rich movement patterns that are essential for understanding vessel behaviors. Annotating semantics for VTs aids in analyzing these behaviors and supports maritime safety management. However, existing methods often focus on either individual or interactive behaviors in isolation, failing to address both simultaneously. Additionally, they are not optimized for large VT datasets and struggle to extract meaningful semantic information effectively. To address these challenges, this paper proposes a method called Semantic Embedding-Enhanced Topic Modeling for Vessel Behaviors (SeeToM). SeeToM integrates the semantics of both single and multiple vessels to construct VT documents. Subsequently, it employs an initialized topic dispersal method to select initial topic centers, thereby improving topic coverage and making SeeToM more suitable for large-scale trajectory data. Furthermore, SeeToM incorporates neural networks into the topic model using Weibull distribution and bidirectional transfer loss, enabling deeper exploration of potential semantic information. Experiments conducted on real datasets from the Zhoushan Archipelago and Sanya demonstrate that SeeToM can effectively extract various vessel semantics, such as acceleration, turn, and collision avoidance. Compared to state-of-the-art baselines, our method achieves improvements of 3.3% in Purity and 3.5% in Normalized Mutual Information (NMI).</p>