Learning Word Embeddings Using Spatial Information

Hideaki Joko, Ryunosuke Oka, Hayato Uchide, Hiroyasu Itsui, Takahiro Otsuka

Abstract: This study proposes a word embedding learning method for intelligent document search which can consider the meanings of place-related named entity word. In the previous method, the word embeddings were learned based on the context in which the words emerged. However, because the previous method missed the spatial information, the word embeddings that share similar contexts (e.g., universities) become similar even though the entities of the words were different (e.g., “” (The University of Tokyo) and “” (Kyoto University)). In the proposed method, the word embeddings are learned based on the spatial information data, which comprise both the object names and coordinates. Therefore, even if the words share similar contexts, the proposed method can learn different word embeddings for different words if those entities are different. The proposed method is evaluated using the synonym search task. As a result, it was observed that the mean reciprocal rank (MRR) for the evaluation data improved using the proposed method in comparison with the previous method. Furthermore, the proposed method improved the MRR by 177% maximally, i.e., from 0.151 to 0.418, in comparison with the previous method.