Improving Ranking in Document based Search Systems

Abstract

In this 21st century, where technology is blooming producing tons of data, efficient retrieval techniques are required to manage these loads of data to endow users with the right information. This paper discusses two neural network techniques applied towards ranking in document-based search systems on two distinct scales: semantic similarity and relevance factor. Semantic similarity focuses on retrieving most contextually similar documents based on a query. Experiments using a semantic approach provides information about how well the system can identify word order. Also, in ideal conditions, the performance was better than traditional benchmark ranking models. The relevance factor focuses on building a neural model based on kernel pooling to work for soft match signals. Experiments are conducted using neural models like K-NRM, Pre-trained embedding models, etc, to prove how their ranking efficiency is better than other traditional models like BM25, RankSVM, DRMM.