TSAI, Richard Tzong-Han, LAI, Yu-Ting, PAI, Pi-Ling, WANG, Yu-Chun, HUANG, Sunny Hui-Min and FAN, I.-Chun, 2017. WeisoEvent: A Ming-Weiso Event Analytics Tool with Named Entity Markup and Spatial-Temporal Information Linking. In: DH. 2017. pdf 
[Accessed 19 March 2024].
In clustering algorithms, each paragraph is represented as a vector. In previous studies, paragraphs have been represented using the vector space model (VSM), which represents each text as a feature vector of terms. However, this approach loses the ordering and ignores semantics. Yet another representation scheme inspired by word2vec is the “Paragraph Vector” proposed by (Le and Mikolov, 2014), an unsupervised framework that learns continuous distributed vectors for pieces of text. In their model, entire paragraphs are represented as vectors. The vector representation is trained to predict the words in a paragraph. More precisely, they concatenate the paragraph vector with several word vectors from a paragraph and predict the following word in the given context. Le’s Paragraph Vector model has many advantages. First, it is mostly unsupervised and works well with sparsely labeled data. Second, it is suitable for text strings of various lengths, ranging from sentences to whole documents. Finally, it can overcome many weaknesses of the bagof-words and bag-of-n-grams models. Because it does not suffer from data scarcity and high dimensionality, it also preserves the ordering and semantic information.
In summary, we propose a classification method which is based on clustering. First, we employ a named entity (NE) recognizer to label texts. Second, we train a paragraph vector model to represent paragraphs as vectors. Third, we cluster paragraphs with length <40 characters. Finally, we use the clustering results as gold-standard categories with which to train a support-vector-machines classifier to predict other paragraphs’ categories.
We compare our method with the state-of-the-art paragraph clustering method using continuous vector space representation proposed by (M. Chinea-Rios et al., 2015). They use word2vec to learn word vectors and represent each sentence by summing the vectors of the words in that sentence. Like Chinea-Rios et al., we use the k-means algorithm to cluster vectors. We set the number of clusters to 68. We refer to the evaluation measures used in (Le and Mikolov, 2014). We generate sets of three paragraphs: two with the same event type and one with a different event type. Each set is referred to as a Paragraph Triplet. The distance between the two vectors with the same event type should be closer than the distance between either of these two and the unrelated one. We collect 923 Paragraph Triplets and compute the accuracy. Our best configuration that combines word dimensions and named entity dimensions to generate paragraph vectors achieves an accuracy of 62.49%, outperforming Chinea-Rios et al.’s pure text-clustering approach (M. Chinea-Rios et al., 2015) by 24.65%.