Abstract

In a multi-domain conversational system, such as Alexa, a key challenge is to enable transfer of actionable information across applications. Most often these systems have evolved independently, with their own local schemas and interoperability requires discovering shared relations between these schemas. There is a body of work dealing with schema matching techniques, and building unified models against a shared ontology. However, this is harder to achieve in a more distributed framework such as Alexa Skills Framework. In this work, I will present research on two different but related work using this theme of leveraging heterogeneous schemas. In the first task, we focus on  semantic parsing. We show that by training the systems jointly across multiple schemas, using a shared intermediate representation,  we can help improve the accuracy of the system even in relatively small data conditions. Furthermore, this improvement exists even when an auxiliary task such as syntactic parsing is used in this multi-task setup. In the second task, we focus on transfer learning from larger models to smaller skill specific models across diverse label sets. We present a novel architecture that can better exploit the correlation between labels and demonstrate how it can be used to improve a named entity classification task.

Biography

Dr. Lambert Mathias is currently a principal scientist working at Alexa Machine Learning, based out of Seattle, where he currently leads efforts on several aspects of conversational AI. In his role at Amazon, Lambert has worked on a variety of NLP driven initiatives, including laying down the foundational technologies that power the language understanding systems in Alexa. Prior to joining Amazon, Lambert worked as a Sr. Research Scientist in the acoustic modeling algorithms group, at Nuance, where he helped develop methods and techniques that helped power the first mobile voice platform behind SIRI. Lambert is an alumni of CLSP; he received his PhD in 2008, where he worked with Dr. Bill Byrne on statistical methods for speech translation.