The amount of records kept by archives is growing in tandem with the speed at which civilization is progressing. Therefore, in order to facilitate the electronic administration of documents, several archive organizations have invested in archive management software. Unfortunately, the automated file categorization feature is absent from the majority of archive management software now used by archive organizations. An automatic pattern identification approach for archival texts that is appropriate for relevant institutions is investigated in this study with the aim of lightening the burden on archive business staff. Recent years have seen the fast development of neural networks as an effective tool for data mining. Hence, a neural network-based approach to archive text categorization is built in this research. One solution to the problem of inadequate feature extraction by individual neural network models is the suggestion of a hybrid model that combines a residual network with a self-attention mechanism. The model begins by extracting local and contextual semantic characteristics from the input text using Bidirectional Gated Recurrent Unit—Convolutional Neural Network (BiGRU-CNN). After the convolution procedure, the data is weighted using the self-attention technique to remove important text characteristics. Lastly, a residual network is utilized to tackle the problems of gradient disappearance and network deterioration in deep neural networks. This is done after normalizing the pooled feature information. The model calculation procedure also makes use of the smoother Mish function as a non-linear activation function to enhance the generalizability of the model.