基于分布的中文词表示研究

针对基于分布的中文词表示构造过程中的参数选择问题进行了系统性的研究。选择了六种参数进行对比实验，在中文语义相似度任务上对不同参数设置下得到的中文词表示的质量进行了评估。实验结果表明，通过选择合适的参数，基于分布的词表示在中文语义相似度任务上能够得到较高的性能，而且，这种高维的词分布表示的质量甚至优于目前流行的基于神经网络（Skip-gram）或矩阵分解（GloVe）得到的低维的词表示。

Study of distributional representation of Chinese words

To solve the problem of parameters selection in the process of constructing the distributional representations of Chinese words, this paper performed a systematic study. Six kinds of parameters were selected for comparison experiments, and the quality of the distributional representations of Chinese words obtained under different parameter settings was evaluated on the Chinese semantic similarity task. The experimental results show that, by choosing appropriate parameters, the distributional representations of Chinese words can also get higher performance on the similarity task, Moreover, the quality of such high-dimensional distributional representations is even superior to low-dimensional word representations based on neural network or matrix factorization.