改进的HMM模型在特征抽取上的应用

MHMM Used in Feature Extracting

摘要:

目前，对情感分类常用的特征抽取方法是基于词典的向量空间模型（VSM），潜在的语义分析（LSA）和基于无监督算法的词嵌入（word2vec），和随机词向量法，这些方法都是对单个词语进行处理。本文通过哈工大词云对采集的豆瓣评论数据集进行语义角色进行的标记以后，采用了改进的隐马尔科夫模型（MHMM）对词对向量进行特征构建，并将其作为一个序列片段作为长短记忆门（LSTM）的输入，最后使用softmax函数对动态循环神经网络输出的序列进行分类，本次实验使用了交叉熵作为优化函数，采用了随机梯度下降法对优化函数进行迭代产生最优解，实验结果证明了该方法对豆瓣影评数据进行情感分类产生了更好的效果。

Abstract:

Nowadays, In the process of emotion analysis, the most commonly used feature method is based on the dictionary vector space model (VSM), latent semantic analysis (LSA) and word2vec which is based on unsupervised algorithm. All of the above models are applied in a single word. In this paper, having remarked the Douban datasets through the semantic role which was crawled on the internet, we making use of the result to traverse the order tree, and taking advantage of Modified Hidden Markov Model, and in the experiment the optimization function is cross entropy and the way that we find the best solution is stochastic gradient decline. In the end, we find LSTM model for test and verify the accuracy of the feature extraction, it turn out to be that this mean has a good effect on sentiment analysis.