基于LSTM/GCN的在线学习文本特征提取方法

Online Learning Text Feature Extraction Method Based on LSTM/GCN

摘要: 传统方法对在线学习文本进行特征筛选往往费时费力且迁移性较差。针对这一问题，根据在线学习文本短，专业词汇多，文本间结构信息丰富等特点，提出基于LSTM/GCN对Doc2Vec所得文本向量中文本–文本关系进行强化的文本嵌入方法，以解决传统方法中文本在投影到嵌入空间后结构信息丢失的问题。并提出指标MeanRank用于量化文本向量中结构信息的留存情况。实验结果表明，方法在指标MeanRank和文本分类精度上优于传统方法。可视化结果表明，增加结构向量使得文本向量在课程内部具有一致连贯性，在课程间更有区分度。

Abstract: Traditional methods for feature filtering of online learning text are often time-consuming and poorly migratory. To address this problem, and based on the characteristics of short texts of online learning text , many specialized vocabularies, and rich structural information between text, an end-to-end text feature extraction method is proposed. The method emphasizes the text-text relationship based on LSTM/GCN by obtaining the text vector based on Doc2Vec model to solve the phenomenon that the traditional method text loses structural information after projection to the embedding space. And the metric MeanRank is proposed to quantify the retention of structural information in the text vector. Experimental results on the Yale Open Course dataset show that the method outperforms traditional methods in terms of metrics MeanRank and text classification accuracy. Visualization of t-distributed stochastic neighbor embedding of text vectors shows that adding structural vectors makes text vectors consistently coherent within courses and more dis-criminative between courses.