基于Fasttext和多融合特征的文本分类模型  

Text Classification Model Based on Fasttext and Multi-Feature Fusion

作　　者：张焱博 郭凯[1] ZHANG Yan-bo;GUO Kai(Beijing University of Post and Telecommunication,Beijing 102206,China)

机构地区：[1]北京邮电大学,北京102206

出　　处：《计算机仿真》2021年第7期461-466,共6页Computer Simulation

摘　　要：针对传统文本分类办法无法有效应对长文本快速收敛的问题,提出了一种基于LSTM-CNN-ATTENTION机制FASTTEXT的文本分类模型。该模型使用预训练词向量将文本信息转换为词向量,通过将词向量送入CNN层、Bi-LSTM层,获得所对应的深度词向量特征,并通过Attention机制使CNN层特征与Bi-LSTM层特征交互,得到融合特征表示。同时,通过将词向量送入FASTTEXT层得到文本信息浅层表示,并与Attention机制的深层特征相拼接,将拼接后的特征映射到分类条目中实现文本分类。实验结果显示,与CNN、Bi-LSTM、AT-LSTM-CNN模型相比,该方法在取得了良好的分类效果前提下,有效地加快了模型学习速度。Aiming at the problem that traditional text classification methods cannot effectively cope with the rapid convergence of long texts,a text classification model based on LSTM-CNN-ATTENTION mechanism-FASTTEXT is proposed.This model uses word2vec to convert text information into word vectors.By sending the word vectors to the CNN layer and the Bi-LSTM layer,the corresponding deep word vector features were obtained.Through attention mechanism,CNN layer features interacted with the Bi-LSTM layer to obtain fusion feature representation.Meanwhile,the word vector was sent to the FASTTEXT layer to obtain a shallow representation of text information,and the deep features of the attention mechanism were spliced together to achieve text classification.The experimental results show that compared with CNN,Bi-LSTM and AT-LSTM-CNN models,this method effectively accelerates the model learning speed under the premise of achieving good classification results.

关 键 词：文本分类 深度学习 特征融合 注意力机制 双向循环神经网络模型 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]