Query-based Single Document Summarization using Hybrid Semantic and Graph-based Approach

In this era of digitization, a plethora of information, spanning a multitude of themes, is available worldwide. The demand for query focused text summarization is growing rapidly due to the presence of irrelevant and repetitive information. Many researchers have immensely studied various unsupervised and supervised techniques for generic text summarization. However, the field of query-based summarization still demands significant research and development. This paper presents a novel hybrid methodology to extract various sentences that are relevant to a user-given question or query. The proposed model uses unsupervised algorithms to first detect the topics in the data by generating an intuitive semantic structure. It then employs a conditional semantic similarity measure to retrieve data containing keywords. An enhanced graph-based approach has been used to include additional query-relevant data as well as discard the redundant information to generate an efficient output summary using a weighted graph. The performance of the model is measured using two evaluation metrics, namely ROUGE-N/L scores and a peer-review survey. The efficiency of the semantic model is evaluated through coherence measure. DUC-2005 and an email dataset have been used for evaluation. The results showcase an improved performance compared to other traditional methods and encourage future work in this field.