The Evaluation of Semantic Similarity Between Phrases for Text Message Classification

Short text or single phrase classification task exists in client inquiries online routing systems and in intelligent conversational systems to identify client intents. One of the crucial problems in this task is the identification and matching of paraphrases. The same intent can be explained by several different phrases, even with completely different words. This problem dramatically complicates the classification task and makes conventional methods of text classification based on word frequency useless. In this work, we propose using probabilistic topic modeling with partial supervised learning and automatic detection of the informative word collocations together with neural vector interpretation of words to evaluate semantic similarity of phrases. These methods was tested on a corpus consisting of question pairs with binary indication and was compared with the Word2Vec and FastText models.