With the rapid development of data technology, a large amount of complex and dynamic text information is generated every day around the world. How to efficiently extract and utilize potential knowledge from it has become an important research direction in the field of information processing. As a key technology to reveal the deep logical relationship between events and infer future trends, event prediction faces higher application difficulty and research value in complex M&A scenarios. However, existing methods have obvious limitations in large-scale data dependence and model design. This paper combines the deep learning framework with the attention mechanism to propose an innovative complex M&A event prediction model (AttDN) to improve the model's adaptability and prediction accuracy in limited labeled data scenarios. In the event recognition stage, this paper designs a small sample learning framework, which effectively improves the accuracy and stability of recognition by accurately modeling the key information points of M&A events and optimizing the feature capture method; in the event prediction stage, this paper abstracts the M&A event chain into a graph structure model, uses the improved neural network framework to model the deep semantics of the complex relationship between events, and combines the attention allocation strategy based on the convolutional network to dynamically weight the key events, thereby significantly enhancing the model's reasoning ability and prediction effect. Through experimental verification on a public M&A event dataset, the model proposed in this paper shows superior performance and significant innovation in complex scenarios. The research results not only provide theoretical and practical support for the dynamic analysis of complex M&A events, but also open up new ideas for the further development of information mining.