A Comparative Analysis of Sentence Embedding Techniques for Document Ranking

Due to the exponential increase in the information on the web, extracting relevant documents for users in a reasonable time becomes a cumbersome task. Also, when user feedback is scarce or unavailable, content-based approaches to extract and rank relevant documents are critical as they suffer from the problem of determining semantic similarity between texts of user queries and documents. Various sentence embedding models exist today that acquire deep semantic representations through training on a large corpus, with the goal of providing transfer learning to a broad range of natural language processing tasks such as document similarity, text summarization, text classification, sentiment analysis, etc. So, in this paper, a comparative analysis of six pretrained sentence embedding techniques has been done to identify the best model suited for document ranking in IR systems. These are SentenceBERT, Universal Sentence Encoder, InferSent, ELMo, XLNet, and Doc2Vec. Four standard datasets CACM, CISI, ADI, and Medline are used to perform all the experiments. It is found that Universal Sentence Encoder and SentenceBERT outperform other techniques on all four datasets in terms of MAP, recall, F-measure, and NDCG. This comparative analysis offers a synthesis of existing work as a single point of entry for practitioners who seek to use pretrained sentence embedding models for document ranking and for scholars who wish to undertake work in a similar domain. The work can be expanded in many directions in the future as various researchers can combine these strategies to build a hybrid document ranking system or query reformulation system in IR.