Enhancing Accuracy of Semantic Relatedness Measurement by Word Single-Meaning Embeddings

We propose a lightweight algorithm of learning word single-meaning embeddings (WSME), by exploring WordNet synsets and Doc2vec document embeddings, to enhance the accuracy of semantic relatedness measurement. In our model, each polyseme is decomposed into a series of monosemous words with diverse WordNet synset tags which represent different word meanings, and there is a one-to-one correspondence between a word meaning and a vector. Our algorithm proceeds in 3 steps. First, the word sense disambiguation of each polyseme in different contexts is achieved by computing the maximum relatedness between the context of this polyseme and all its candidate meaning definitions in WordNet. Second, each tagged word is lemmatized according to its synset tag to alleviate the word sparsity problem caused by polysemes decomposition. Third, the word single-meaning embeddings are learned from the meaning-tagged corpus, and the semantic relatedness between words can be more accurately measured based on such embeddings. Our experimental results show that our algorithm achieves better performance on the semantic relatedness measurement compared with existing techniques.