DeepPurple: Estimating Sentence Semantic Similarity using N-gram Regression Models and Web Snippets

We estimate the semantic similarity between two sentences using regression models with features: 1) n-gram hit rates (lexical matches) between sentences, 2) lexical semantic similarity between non-matching words, and 3) sentence length. Lexical semantic similarity is computed via co-occurrence counts on a corpus harvested from the web using a modified mutual information metric. State-of-the-art results are obtained for semantic similarity computation at the word level, however, the fusion of this information at the sentence level provides only moderate improvement on Task 6 of SemEval'12. Despite the simple features used, regression models provide good performance, especially for shorter sentences, reaching correlation of 0.62 on the SemEval test set.