Next-event prediction in soccer is gaining attention as a core task, with applications in action value estimation, counterfactual simulation, and player style analysis. Prior work, such as the Large Event Model (LEM) framework, has focused on modeling sequences of events in an autoregressive manner. However, these models often overlook the influence of team and player identities. In this work, we extend the LEM framework to incorporate team-level and player-level information for predicting the next action type. Using the Wyscout dataset, we examine the impact of team identity, player identity, and player role on model performance. To this end, we conduct an experiment in which a variant of the LEM model based on a multi-layer perceptron (MLP) is trained, with player and team embeddings included as optional components. We compare the performance of the resulting model variants with heuristic baselines and an XGBoost classifier. Our findings show that, for soccer action type prediction, XGBoost and MLP-based models perform similarly and outperform baseline heuristics. We also confirm that incorporating a longer history of past events improves prediction performance compared to using only one past event. Finally, although including team or player embeddings does not substantially enhance predictive performance, our exploratory data analysis reveals that the learned player embeddings capture meaningful patterns related to player roles and team-specific playing style.