Revisiting Skip-Gram Negative Sampling Model with Regularization
We revisit skip-gram negative sampling (SGNS), a popular neural-network based approach to learning distributed word representation. We first point out the ambiguity issue undermining the SGNS model, in the sense that the word vectors can be entirely distorted without changing the objective value. To resolve this issue, we rectify the SGNS model with quadratic regularization. A theoretical justification, which provides a novel insight into quadratic regularization, is presented. Preliminary experiments are also conducted on Google's analytical reasoning task to support the modified SGNS model.
READ FULL TEXT