【摘要】 中文短文本具有长度短以及上下文依赖强的特点。目前主流的基于词向量的双向循环神经网络分类方法依赖于词向量语义表达以及上下文特征提取能力，其分类准确率有待进一步改善。针对此问题，论文提出融合语义增强的中文短文本分类方法，该方法提出融合语义增强的思想，在词向量表示阶段，引入Bert生成融合字、文本以及位置的向量作为训练文本的词表征进行文本语义增强，接着输送到Bi-GRU网络中提取上下文关系特征，并通过多头注意力机制调整权值强化重要特征表达，最后使用softmax分类器进行文本分类。通过与其他主流方法进行对比论证，实验表明，论文提出的方法在短文本分类效果上有显著提升。更多还原

【Abstract】 Chinese short texts are short in length and highly context-dependent. The current mainstream word vector based bidirectional recurrent neural network classification method relies on word vector semantic expression and context feature extraction capabilities,and its classification accuracy needs to be further improved. This paper proposes a classification method of Chinese short text with semantic enhancement. In the stage of word vector representation,Bert is introduced to generate semantic vectors,which fuse words,text and position vectors as word representations of training text to enhance semantics. Then,context features are extracted from BI-GRU network,and important feature expressions are enhanced by multi-attention mechanism to adjust weights.Finally,it is classified by softmax classifier. By comparison with other mainstream methods,the experiment shows that the method proposed in the paper has a significant improvement in the classification effect.更多还原