'Relevant' Machine Translation with Learning to Rank
Suneel Marthi • Back to Haystack 2019
Learning to Rank (LTR) has been used successfully in several areas of Information Retrieval and Search for constructing ranking models. One of the less explored applications of LTR is in the area of Machine Translation. Machine Translation is important when having to cater to different geographies and locales for news or eCommerce website content. Machine Translation systems often need to handle a large volume of concurrent translation requests from multiple sources in multiple languages.
Typical machine translation involves generating several possible candidate translations and picking the best candidate using an algorithm like BeamSearch. With Deep Learning becoming more ubiquitous in recent times and with Neural Machine Translation(NMT) having replaced the traditional statistical translation methods, it is but natural that we use Deep Learning techniques to rank and pick the best translation.
In this talk, we'll be looking at leveraging TensorFlow's Learning to Rank plugin to pick the most 'Relevant' translation from several possible candidate translations and dig deeper into using Deep Learning techniques for Learning to Rank. The talk will cover details about TensorFlow's LTR plugin, how to train a LTR ranking model for NMT, and the advantages of using LTR for NMT over regular NMT and how LTR compensates for some of the shortcomings with regular NMT by using the 'Context' of individual sentences in choosing the best translation.