Abstract: A global word vector training model based on pointwise mutual information was presented. The model used another correlation information, the ratio of the joint probability and the product of the marginal probability, to depict the relationship between words and avoid the shortcoming of conditional probability. In order to verify the validity of our model, we trained word embedding by GloVe and our model in the same situation and then carried out experiments with word analogy and word similarity separately using these two word embeddings. Experiments showed our model has achieved 10.50%, 4.43%, 1.02% better accuracy rate than the GloVe model does in sematic experiments of word analogy at 100 dimensionality, 200 dimensionality, 300 dimensionality respectively. Accuracy rate has also gained 5%-6% rise in word similarity experiments. The results show that our model can capture semantics of words more effectively.

Key words: pointwise mutual information, word vector, GloVe