Abstract: 【Purpose/significance】This paper tries to find an effective way to classify the non-structured text, aiming to improve the ef?
ficiency of corporate problem solving and provides decision advice to enterprise users.【Method/process】In order to solve the problems of word vector feature embedding, such as unavailability of polysemous words, sparse features and difficulty in feature extraction, this paper proposes a multi-channel hierarchical feature text classification model based on sentence features (SFM-DCNN).First, the mod? el upgrades feature embedding from traditional word feature embedding to sentence feature embedding through Bert sentence vector modeling, effectively obtaining semantic features such as word polysemy, word position and inter-word association. Secondly, the em? bedded features from multiple levels for feature reinforcement to obtain features closer to the original semantics by constructing a multi-channel deep convolution model.【Result/conclusion】We examined the proposed model with three different texts and found its precision is much higher than the traditional methods, indicating that the model has certain reference significance.【Innovation/limita?tion】Based on the problem of polysemy and feature sparsity in text classification, Bert is used to extract global semantic informationand multi-channel deep convolution is used to obtain local hierarchical features, however, due to time and equipment conditions, the model has not been further trained and the experimental data set is not sufficient.