Visit our website: http://bit.ly/2GtXaiw
Multi Lingual Label Printing – An optimized AI Approach From Translation to Vector Space Allocation for Multilingual Label Generation – A Real-World Case Study
Shivpratap Singh, Senior Advisor, CVS Caremark
Abstract:
As per the multilingual drug label law enforcement policy each pharmacy chain should be able to provide the drug label on the bottles in 32 dispensing Languages . This bring us a challenge of providing a user friendly experience to all customers. Also there were performance challenges in scaling our current system where we had challenge to generate 72 labels/hours and were able to achieve ~30/hours.
Problem get intensified as we keep on getting request to support new set of languages. We did plan to disrupt our existing model and evolve a new system to move label generation engine from a language-specific models to multilingual embedding’s that serve as a part of our Unified data platform strategy. This scenario currently deal with two types :NLP Syntactic Parsing and Machine language translation. One important task in NLP is text classification. Since we have to handle these text in multiple languages so we recommending to make text classification ‘multi lingual’ to develop multi lingual word embedding .With this technique embedding for every language does exist in the same vector space and maintain the property that words with similar meaning are close together in vector space.
Approach to build a multi lingual classifier and to switch over translation based engine to a vector space classifier model exponentially helps us scale to more languages and improve the overall efficiency/throughput from 32/hour to 78/hour on our retail pharmacy chain
Add comment