the

Although both topic distributions and word embeddings are used to capture the semantic characteristics of user-generated text, they have different focuses. Topic distributions are more suitable to extract topical themes\ from text based on word co-occurrence patterns \(essentially taking the whole document as the context window\) while word embeddings are more suitable to capture the semantics between words from local context windows, usually comprising 3 words before an\ d after the target word. Hence, we keep both types of text features in our approach. It is worth noting that our method is a tree-based approach, which can effectively handle information redundancy, i.e., if a feature contains redundant information given \ the tree that is being constructed, it will be pushed to a lower rank during attribute selection.