A Method for Visual Interpretation of Word2Vec Static Vector Space
Abstract:
Starting its creation, static vector word models was able solving the task of finding word analogies. This task finds a vector of parallel transition that reflects a word into another one changing a property of a former word. Since the task of finding word analogies lacks of precision, researchers are trying to investigate peculiarities of vector models and corresponding semantic spaces. In this paper, we introduce a new method of visual interpretation of such models. Our method uses thematic word collections, separation of semantic space into word groups using LSA method, and visualization of results using heatmap. We found out that vector semantic space can be interpreted not only on a local level, but on the global one as well. In the former case, the resulting grouping of words depends on the text collection used for the model training. Our method is applicable for finding several top layers because of exponential decreasing of the size of a group.
