Comparing the Performance of Neural and Statistical Sentence Embeddings on Summarization and Word Sense Disambiguation

We analyzed the performance of two sentence embeddings: SIF (Smoothed Inverse Frequency) created using weighted GloVe word embeddings, and sent2vec, trained using a neural network. Using these sentence embeddings without modification, we compared and contrasted their performance on extractive text summarization and word sense disambiguation using existing methods tailored for sentence embeddings. We find that our results are better than the simplest baselines and approach competitive baselines for both these tasks, proving that sentence embeddings are to some extent successful in capturing the structure of language.