Deep Neural Query Understanding System at Expedia Group

摘要

Understanding customer intent expressed through search queries is necessary to not only provide the best shopping experience to Expedia Group customers but also to maximize marketing returns. Natural language Understanding (NLU) has ubiquitous commercial application in search, conversational platforms and more. Search queries are notoriously terse, noisy and lack grammatical cues making NLU a challenging task. Multi-lingual market scalability - a highly desirable feature for global travel agent - further add complexity. In this work, we present our NLU System for such search queries in the travel domain using multi-lingual deep learning models that perform these broad tasks: intent classification, named entity recognition and linking. We propose an alternate framework that significantly improves recognition and resolution of ill-defined sparse entities. Our system also includes cross-lingual transfer learning components featuring active learning loop to scale these models to multiple languages with minimal but high quality annotation by localization experts. We explain the business problem these models address, idiosyncrasies of our data, architecture details and implementation trade-offs.