results. Some researchers consider the reliability of basic clustering, but consider a
basic clustering as an individual without considering each cluster [9, 11, 12]. In this
paper, Gini coefficient is used to evaluate the reliability of base clustering, and hierarchical clustering algorithm is used to integrate the final clustering.
To solve the above problems, this paper proposes a short text ensemble clustering
method based on CNN. Firstly, the high-dimensional representation of word vectors is
trained by word2vec method. Then, the feature vectors are extracted by CNN, and the
lower-dimensional representation of word vectors is obtained. Next, the feature vectors
are clustered by some traditional clustering methods. At last, the clustering ensemble
algorithm proposed in this paper is used for clustering ensemble.
2 Vectorization of Short Text
2.1 Pre-processing Short Texts
For the Chinese text, the pretreatment process mainly includes the removal of punctuation marks, word segmentation, and removal of stop words. Punctuation is removed
for subsequent word segmentation of the remaining plain text where word segmentation refers to the segmentation of a continuous Chinese text into separate entries. In
Chinese texts, there are many words with high frequency but no specific meanings,
such as “的” and “在,” which are called as stop words. The purpose of eliminating stop
words is to preserve the core words as much as possible so that the subsequent generated word vectors are not affected by noise and improve the quality of word vectors.
2.2 Generating Word Vectors
The so-called word embedding refers to the technology mapping words into a new
space and expressing them in the form of vectors. Traditional word vector representations, such as one-hot, are sparse and multifarious, which can cause dimension disasters and ineffective access to semantic information in practical use. Mikolov et al.
proposed word2vec model which is based on neural network [13]. Word2vec model
trains the language model through the context and context semantic relations of words
and obtains word vector as the auxiliary product in the training process.
The word2vec has two important models—CBOW model and Skip-gram model.
CBOW model uses context to predict center word, while Skip-gram model uses center
word to predict context. This paper chooses CBOW model to train to get the vector
representation of k-dimension of each word and fills all short texts according to the
maximum length N in short texts, so that all texts are presented in the matrix form of
n * K. The uniform format is also convenient for subsequent batch convolution
operations.
2.3 Extracting Feature Vectors
The structure of neural network used to process short text in this paper includes input
layer, convolution layer, max-pooling layer, and full connection layer.
Research on Chinese Short Text Clustering …
623
basic clustering as an individual without considering each cluster [9, 11, 12]. In this
paper, Gini coefficient is used to evaluate the reliability of base clustering, and hierarchical clustering algorithm is used to integrate the final clustering.
To solve the above problems, this paper proposes a short text ensemble clustering
method based on CNN. Firstly, the high-dimensional representation of word vectors is
trained by word2vec method. Then, the feature vectors are extracted by CNN, and the
lower-dimensional representation of word vectors is obtained. Next, the feature vectors
are clustered by some traditional clustering methods. At last, the clustering ensemble
algorithm proposed in this paper is used for clustering ensemble.
2 Vectorization of Short Text
2.1 Pre-processing Short Texts
For the Chinese text, the pretreatment process mainly includes the removal of punctuation marks, word segmentation, and removal of stop words. Punctuation is removed
for subsequent word segmentation of the remaining plain text where word segmentation refers to the segmentation of a continuous Chinese text into separate entries. In
Chinese texts, there are many words with high frequency but no specific meanings,
such as “的” and “在,” which are called as stop words. The purpose of eliminating stop
words is to preserve the core words as much as possible so that the subsequent generated word vectors are not affected by noise and improve the quality of word vectors.
2.2 Generating Word Vectors
The so-called word embedding refers to the technology mapping words into a new
space and expressing them in the form of vectors. Traditional word vector representations, such as one-hot, are sparse and multifarious, which can cause dimension disasters and ineffective access to semantic information in practical use. Mikolov et al.
proposed word2vec model which is based on neural network [13]. Word2vec model
trains the language model through the context and context semantic relations of words
and obtains word vector as the auxiliary product in the training process.
The word2vec has two important models—CBOW model and Skip-gram model.
CBOW model uses context to predict center word, while Skip-gram model uses center
word to predict context. This paper chooses CBOW model to train to get the vector
representation of k-dimension of each word and fills all short texts according to the
maximum length N in short texts, so that all texts are presented in the matrix form of
n * K. The uniform format is also convenient for subsequent batch convolution
operations.
2.3 Extracting Feature Vectors
The structure of neural network used to process short text in this paper includes input
layer, convolution layer, max-pooling layer, and full connection layer.
Research on Chinese Short Text Clustering …
623