Improving Word Representation Quality Trained by word2vec via a More Efficient Hierarchical Clustering Method
Abstract: In traditional word2vec methods, hierarchical softmax algorithm uses the whole vocabulary to construct a Huffman tree and it trains each pair of words just in logarithmic time consumption. But due to the lack of consideration about cooperation of each word in the corpus, it will reduce the performance of language model and the trained word vectors. In this paper, we substitute a purely data-driven method for the original Huffman-tree method to rebuild the binary tree. The new construction method utilizes the semantical and syntactical cooperation of words to cluster the words hierarchically. The cooperation of words is reflected in the word vectors which collected from the initial Huffman-tree training procedure. Our methods substantially improve the performances of word vectors in semantical and syntactical tasks.
Loading