Contextual Analysis for the Representation of Words

Building the representation of word meanings is one of the key challenges in many language-based applications from document understanding, text summarization to sentiment analysis. One of the reasons to make this task harder is that word meanings involve not only the word surfaces in contexts, but the human experiences in specific domains. Previous work in the field considers these issues separately by analysing text contents in one hand and dealing with knowledge-based information on the other hand. In this work, we address this issue by accumulating contextual information of words and knowledge-based contents to construct the representation of words. We evaluate the effectiveness of the representation via the task of semantic similarity on standard benchmarks. The experimental results show the strong correlation between the proposed word representation to the perception of human in the task of semantic similarity measure.