Text summary evaluation based on interpretable semantic textual similarity

International audienceText summarization methods are much needed to tackle the ever-increasing volume of text data, accessible online to help us find the relevant information and quicker ingestion of relevant data. In this paper, we have reported two issues related to text summarization. At first, we developed a Deep Neural Network (DNN) based abstractive text summarization method. After that, a chunk alignment-based interpretable semantic textual similarity (iSTS) method is designed to evaluate the quality of the summary text with reference to the main text. We have used an attention-based encoder-decoder Recurrent Neural Network (RNN) model to develop the abstractive text summarization method. The encoder compresses the sequence information into a sequence of vectors to save the important information. A Long Short Term Memory (LSTM) based RNN model, composed of three stacks, takes the words in a sequence and produces the output as hidden states. During training, we have used word embeddings for each word with a dimension of 128. At first, the efficiency of the summary text in reference to the main text is evaluated using the BLEU score and ROUGE matrix. Further, the results are compared with the proposed iSTS-based evaluation measure. The quality of the summary text is accessed based on the similarity score, and for that, we have trained a multivariate supervised linear regression model. The supervised algorithm is trained with the features extracted from a pair of the chunk itself. The string similarity, distributional word representations, and relatedness scores between the chunks are provided as a feature vector to get the similarity score