Document vectorization method using network information of words

We propose a new method for vectorizing a document using the relational characteristics of the words in the document. For the relational characteristics, we use two types of relational information of a word: 1) the centrality measures of the word and 2) the number of times that the word is used with other words in the document. We propose these methods mainly because information regarding the relations of a word to other words in the document are likely to better represent the unique characteristics of the document than the frequency-based methods (e.g., term frequency and term frequency–inverse document frequency). In experiments using a corpus consisting of 14 documents pertaining to four different topics, the results of clustering analysis using cosine similarities between vectors of relational information for words were comparable to (and more accurate than in some cases) those obtained using vectors of frequency-based methods. The clustering analysis using vectors of tie weights between words yielded the most accurate result. Although the results obtained for the small dataset used in this study can hardly be generalized, they suggest that at least in some cases, vectorization of a document using the relational characteristics of the words can provide more accurate results than the frequency-based vectors.