Home | Independent Research | Last | Next |
R. JeffreyPennington and C. Manning. Glove: Global vectors for word representation. 2014.
To better deal with word prepresentations: word analogy, word similarity, and named entity recognition tasks, the paper construct a new model GloVe (for Global Vectors), which is able to capture the global corpus statistics.
The paper combined count-based methods and prediction-based methods for the unsupervised learning of word representations, proposing a new cost function
Where the weighting function f(Xij) looks like
The paper introduced how they derived the function starting from simple co-occurrence probablities. And how it is related to other models such as skip-gram via cross entropy error.
The paper also proved that the complexity of the model is based on hyper-paramter, always smaller than the on-line window-based method.