Semantic Dependency Word Representation

Abstract

We provide a novel model architecture for learning Chinese word representations with semantic dependency information. These representations can solve two problems, the first is to avoid polysemy, the second is to capture remote features. Our representations are learned from a Chinese semantic dependency model and a normal word embedding model (e.g.,CBOW or Skip-gram). Using CBOW or Skip-gram model, the word representations can get some semantic information from the position of words. Semantic dependency model can find different word pairs in the sentence and predict the semantic relationship between each word pair. The first word of each word pair represents the current word, and the second word represents the target word. By fusing the semantic information after semantic analysis into the basic word representations, we can get more accurate and dynamic word representations. More importantly, we have achieved advanced results when we use these word representations in sentiment classification task and subject classification task.