1. Divided sentences into terms, analyzed terms properties and syntactic dependency by LTP-Cloud with Java.

2. Learned word vectors by word2vec for croups

3. Extracted words features like nouns, adjectives, adverbs and verbs, and extracted sentences structurefeatures like ATT, ADV and so on with python. Processed these features considering the word no.

4. Represented corpus by matching word vectors with features, and trained the labeled vector document by SVM Perf to separately do predictions for words features and sentences structure features.

5. Combined the predictions of words features and that of sentences structure features together by learning weight and threshold, which improved accuracy about 15% for only training words features