Measuring the text similarity is a fundamental technology in power intelligent interaction platform. In view of the short informal text which is short in length and does not conform to strict grammar rules in power intelligent interaction platform, a new text similarity estimation method is designed. Instead of depending mainly on unrepresentative features for short text including part-of-speech (PoS), the proposed measurement processes short texts on three structural hierarchies, including words, phrases and sentences. Based on position, reorder cost and generative entropy of lower-level structures in higher-level structures, new features are extracted from structural information of texts on phrase-level and sentence-level. Similarity among short text is estimated with both new features and vectorized representations on word-level. Experimental results on real-world corpus show the practicality and superiority of the proposed similarity measurement.