Research Article
Research on Fund News Classification Method Based on Multi-level Model Fusion
@INPROCEEDINGS{10.1007/978-3-031-71716-1_16, author={Ju-Xiang Hu and Xue-Qiang Lv and Xin-Dong You and Jian-She Zhou}, title={Research on Fund News Classification Method Based on Multi-level Model Fusion}, proceedings={Machine Learning and Intelligent Communication. 8th EAI International Conference, MLICOM 2023, Beijing, China, December 17, 2023, Proceedings}, proceedings_a={MLICOM}, year={2024}, month={9}, keywords={Word location characteristics Attention mechanism Multi-feature Text classification}, doi={10.1007/978-3-031-71716-1_16} }- Ju-Xiang Hu
Xue-Qiang Lv
Xin-Dong You
Jian-She Zhou
Year: 2024
Research on Fund News Classification Method Based on Multi-level Model Fusion
MLICOM
Springer
DOI: 10.1007/978-3-031-71716-1_16
Abstract
Aiming at many challenges in the field of fund news text classification, including the difficulty of capturing context and dealing with unbalanced feature weights by a single deep learning model, this paper proposes a multi-level model fusion based fund news text classification method. Through in-depth analysis of the structure and characteristics of fund news texts, this method constructs an input matrix based on the integration of multi-level models, and then adopts softmax to complete the classification task of fund news texts. The whole model is referred to as PosiTransBiAttention model (that is, the combination of word location embedding, Transformer, BiLSTM and multi-head attention). The unique feature of PosiTransBiAttention model is that it can fully integrate multi-level features, so as to effectively solve the problem of fund news text classification. The experimental results verify the effectiveness of the model in the fund news text classification task, and the accuracy rate reaches 93.85%. The experimental results further demonstrate the significant performance advantages of the model in this field. Through the combination of multi-feature fusion and multi-level model, this method provides a powerful reference and inspiration for solving the text classification problem in similar fields.
