Joint Training Graph Neural Network for the Bidding Project Title Short Text Classification

With the advent of the information era, the demand for data processing speed and scale is far beyond the capability of past manual methods. Therefore, in the face of complex and complicated bidding information, how to select the bidding projects that meet their needs and process them in a shorter time, the classification of project titles has become an urgent problem to be solved. Considering the characteristics of the short text of bidding titles, We propose BESGN by using the large-scale pre-training model BERT to obtain text contextual information, which facilitates learning to generate good representations of the target text. We also construct a bipartite graph structure using graph neural networks for inter-neighborhood node messaging to capture information between different granularities in the text and overcome its heterogeneity. The two are fused to maximize the preservation of contextual information, compensate for the limitations of short texts, and enable automatic labeling and classification of bidding projects. We conduct experiments on bidding and benchmark datasets and compare BESGN with other classification methods. The results show that BESGN outperforms other models in terms of classification accuracy, especially in handling short texts.