现有情感分类模型多采用单一方面词与文本单层粒度的简单匹配策略,或依赖注意力机制实现特征抽取,此类方法普遍存在文本结构信息利用率不足、特征筛选针对性弱的缺陷,导致分析结果粗放化且模型泛化性能受限。针对该问题,本文提出一种分层树结构注意力模型,模型首先将已经特征编码后的字向量通过自编码机组合成:词汇、子句、句子三种粒度的语料,然后利用树状图神经网络对其进行多层递进的权重分配,再通过最长路径算法筛选出关键节点,最后使用由归一化层和卷积层构成的分类网络获得分类结果。经过实验模型在公开数据集上的准确度和F1分数比同类型模型拥有更好的表现。Existing sentiment classification models mostly use a simple matching strategy of single aspect word and text single-layer granularity, or rely on the attention mechanism to achieve feature extraction. Such methods generally have the defects in insufficient utilization of text structure information and weak targeting of feature screening, resulting in sloppy analysis results and limited model generalization performance. To address this problem, this paper proposes a hierarchical tree-structured attention model, which first synthesizes the word vectors that have been feature-encoded into three granularity corpus: vocabulary, clauses, and sentences by self-coding group, then uses a tree-graph neural network to assign weights to them in multiple layers progressively, and then filters out the key nodes by the longest-path algorithm, and finally obtains the classification results by using the classification network consisting of the normalization layer and convolutional layer. classification results. The experimental model has better accuracy and F1 score than the same type of model on the public dataset.
