基于卷积神经网络的多通道特征表示文本分类模型

尽管长短期记忆网络(long short-term memory,LSTM)、卷积神经网络(convolutional neural network,CNN)及其结合体在文本分类任务中取得了很大的突破.但这类模型在对序列信息进行编码时,往往无法同时考虑当前时刻之前和之后的状态,从而导致最后分类效果不佳.此外,多版本预训练词向量比单个版本的预训练词向量包含更多的信息.因此提出了一种基于CNN的多通道特征表示文本分类模型(multi-channel feature representation text classification model based on CNN,MC-CNN).该模型首先通过两个不同的双向长短期记忆(bi-directional long short-term memory,Bi-LSTM)来对不同来源词向量所表示的文本序列进行正逆序上的特征提取,并以此形成多通道特征;然后利用多尺度卷积网络来进一步使得模型能够同时充分考虑到当前时刻之前以及之后的信息,从而更加有效地进行文本分类.MC-CNN在MR、SST-2、TREC、AG、Yelp_F、Yelp_P数据集上分别达到了81.6％、87.4％、98.6％、94.1％、65.9％、96.8％的准确率,实验结果表明本文模型MC-CNN在文本分类任务中具有优异的效果.

Multi-channel feature representation text classification model based on CNN

Although the Long Short-Term Memory (LSTM), the Convolutional Neural Network (CNN) and their combinations have made great breakthroughs in the text classification task. However, when encoding se-quence information, such models often fail to take into account the state before and after the current moment, which results in poor classification effect. In addition, multiple versions of a pre-trained word vector contain more information than a single version of a pre-trained word vector. Therefore, a multi-channel feature representation text classification model based on CNN (MC-CNN) is proposed to solve this problem. The model firstly extracts the features of the text sequence represented by different source word vectors by using two different Bi-LSTM (Bi-directional Long Short-Term Memory) to form multi-channel features. And then, a multi-scale convolutional network further enables the model to take into account the state information before and after the current time. The accuracy of MC-CNN in MR, SST-2, TREC, AG, Yelp_F and Yelp_P datasets was 81.6%, 87.4%, 98.6%, 94.1%, 65.9% and 96.8%, respectively. The experimental results show that the model MC-CNN proposed in this paper has excellent effects in text classification tasks.