Refining Word Embeddings for Funniness Predicting Sentimental Analysis Tasks

Sentiment analysis lies within one of the core fields for natural language processing. Concerning this field, word embeddings have been widely used as an effective way of representing words based on distributional semantics. In this paper, we focus on improving the GloVe embedding used for predicting the funniness of edited news headlines. The GloVe embedding of different dimensions is further tuned with the news headline dataset provided by SemEval, using the Continuous Bag of Words model and Skip-gram model, respectively. A recurrent neural network model is developed as the regression model for prediction task. A set of experiments are designed to evaluate the performance of various refined GloVe embeddings. We found out that original GloVe embedding and refined embeddings share similar performance in our current work. The prediction is more accurate for all refined embeddings when using 100 dimensions rather than 300 dimensions embeddings.