Google Patents Filing Details New Text Classification Method for Office Systems

While traditional office automation approaches historically relied on multi-label document categorization based solely on body text, newly surfaced patent documentation reveals an intentional shift toward incorporating document titles and government approval commentary into classification models. According to Google Patents, prior methods frequently suffered from imperfect classification labels and diminished accuracy because they ignored the structural weight of document headings.

Building Document Word Vectors and Processing Workflows

The methodology outlined in patent document CN117891939A begins with structured preprocessing procedures. As detailed in the filing, construction of the document word vector library involves data acquisition, data cleaning, word segmentation, and part-of-speech tagging. This is paired with label preprocessing, which extracts document tag sets and serializes dataset tag subsets using TF-IDF and Fasttext algorithms.

Underpinning the entire document management environment is a BPM workflow engine. Google Patents notes that this workflow integration allows intelligent office automation systems to adjust and continuously optimize business processes, helping enterprises adapt quickly to changing operational requirements.

Feature Extraction via Self-Inertia Weight Particle Swarm Optimization

Once text and labels are preprocessed, the system extracts document feature sets using a self-inertia weight self-adaptive particle swarm algorithm, identified in the text as SIW-APSO. In this algorithm, individual particles continuously search for coordinate points in space to determine subsequent displacement.

The inertial part, which consists of inertia weight and current particle displacement, representing the degree of dependence on the current motion state.
The best individual position reached by the particle.
The best position achieved by the entire group.

Furthermore, a full connection layer reduces article information and title information extracted by the word coding layer down to the same dimension as the original word vector.

Convolutional Neural Network Classification and Document Flow Recommendations

Following feature extraction, document information identification relies on a convolutional neural network model. According to Google Patents, the CNN model structure incorporates an input layer, multiple convolutional layers, a pooling layer, a full-connection layer, and an output softmax layer, with each convolution layer containing different convolution kernels denoted as ω.

To handle complex government document flows and leading approval comments, the system constructs vector representations based on BPM workflow information.

The intelligent document recommendation and the importance of the government document flow complexity and the leading approval comments are fully considered, part of document flow is completely according to the characteristics of the leading approval comments, document BPM workflow information construction vectors are constructed, the document approval comments are constructed on the basis of the ordering and positions of users, the government document characteristics are met, and the recommendation accuracy is improved.
Google Patents, CN117891939A

This approach ensures that the system effectively leverages workflow data and user positioning to optimize the overall efficiency of government document recommendations.