A Comparison of RNN based Seq2Seq and Transformer Models for Detecting Similar Questions

There are millions and billions of questions on various knowledge sharing websites like Quora.It’s very common that various questions are duplicate in it. An automatic tool for detectingsimilar questions out of these millions of questions is known as similar question detectionapproach. So, this paper presents an approach for detecting similar questions. The semanticsimilarity between pairs of questions represents that questions are similar. The approach detectssimilar questions by comparing the sentence embeddings. All the questions from Quoraquestion pair dataset have been converted into sentence embeddings and then the givenquestion is passed to the model for detecting similar questions. The comparison of detectingsimilar questions has been done with two methods i.e. Cosine and Jaccard similarity. Thecombination of Jaccard and cosine similarity can perform better for detecting similar questionsas Jaccard similarity is able to detect lexical and cosine can detect semantic similarity. Theimportant phase in this study is the conversion of questions into embeddings. So, the questionshave been converted into embeddings using three models: average vector, RNN based Seq2Seqand transformer model. Two datasets have been used for the comparison: one is Quora questionpairs dataset in English and second is the translated Quora question pairs in Punjabi. The BLEUsimilarity metric has been applied for the evaluation of the proposed similar question detectionapproach. The embeddings of the detected and the input question have also been compared forin-depth evaluation.