Using Syntactic Relationships to Classify Construction Documents Based on Machine Learning Methods

The paper regards the possibility of improving the numerical embeddings provided to the different models for text classification. Most ways to produce embeddings are based on the frequency approach. As a result, they don't take into account the semantic meaning of each token. The author shows that resorting to the syntactic relationships of a sentence and the corresponding modification of input embeddings based on TF-IDF can improve the result of classifying construction document texts to a certain extent. Multiplying the adjacency matrix representing syntactic relationships in a sentence by the feature matrix may result in a significant improvement of embeddings for text classification. Neural networks and the method of nearest neighbors (KNN) are used to evaluate the obtained results.