Text classification is an issue of high priority in text mining, information retrieval that needs to address the problem of capturing the semantic information of the text. However, several approaches are used to detect the similarity in short sentences, most of these miss the semantic information. This paper introduces a hybrid framework to classify semantically similar short texts from a given set of documents. A real-life dataset – Quora Question Pairs is used for this purpose. In the proposed framework, the question pairs of short texts are pre-processed to eliminate junk information and 25 tokens, and string-equivalence features are engineered from the dataset, which plays a major role in classification. The redundant and overlapping features are removed and word vectors are created by using TF-IDF weighted average FastText approach. A 623-dimensional data model is obtained combining all the obtained features, and the same is then fed to the Light Gradient Boosting Machine for classification. At last, the hyperparameters are tuned to attain optimized log_loss. The experimental results show that the proposed framework can achieve 81.47% accuracy which is at par with the other state-of-art models.
