Current conversational agents and chatbots are limited in their functionality, as they have to be explicitly trained via sample user sentences (utterances) that are mapped to customer defined actions (intents). Given the variations on how humans might ask a given question, it’s not feasible to provide all possible utterances, limiting the conversational agent’s ability to understand and interpret user inputs. For example, if I ask Alexa, “Is it cold tomorrow?” it correctly understands the utterance and maps it to the weather intent and tells me the weather; however if I ask, “Do I need to wear my sweater tomorrow?” it can’t correctly interpret the utterance. Similarly, robotic process agents (RPA) are great for predefined rule-based tasks, actual interactions with a user don’t always follow a script.
Recent advances in deep learning frameworks for NLP, including ULMFiT, ELMo, and Open AI transformer, allow models to go beyond the shallow representations provided by word vectors to provide deep hierarchical representations of language, allowing the capture of high-level semantic concepts. Just like the ImageNet challenge allowed the emergence of generalizable model features for image classification across different categories, these frameworks enable the creation of models that learn the higher-level nuances of language, vastly improving natural language understanding (NLU). This advance in NLU allows a conversational agent to leverage pretrained models to understand that “wearing a sweater” is associated with cold weather and ask a clarifying question, “Do you want to check the weather for tomorrow?"
Sumeet Vij and Matt Speck showcase an innovative application of deep learning to power cognitive conversational agents. You’ll learn how chatbots can overcome the limitations of limited training datasets by leveraging transfer learning and deep pretrained models for NLP and how machine learning can advance robotic process automation (RPA) from “robotic” to “cognitive” automation.
Sumeet Vij is a director in the Strategic Innovation Group (SIG) at Booz Allen Hamilton, where he leads multiple client engagements, research, and strategic partnerships in the field of AI, digital personalization, recommendation systems, chatbots, digital assistants, and conversational commerce. Sumeet is also the practice lead for next-generation digital experiences powered by AI and data science, helping with the large-scale analysis of data and its use to quickly provide deeper insights, create new capabilities, and drive down costs.
Matt Speck is a data engineer and senior consultant in the Strategic Innovation Group (SIG) at Booz Allen Hamilton, where he works on cognitive solutions projects, building intelligent search and chat applications. Previously, he taught data science and Python at General Assembly, a coding boot camp with locations across the globe.
Comments on this page are now closed.
For exhibition and sponsorship opportunities, email email@example.com
For information on trade opportunities with O'Reilly conferences, email firstname.lastname@example.org
View a complete list of AI contacts
©2019, O'Reilly Media, Inc. • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • email@example.com
I am very excited to attend this conference. Almost all topics are related to my field of research especially NLP/NLU.
In this section “Building Intelligent Conversational Agents with Transfer Learning and Cognitive Automation”, I have a quick question, does this topic covers handling multi-language in single sentence of user response. If so, how is sentence embedding created. And, what will be the Bot/Agent response (in terms of language). Also, currently, I am working with BERT published by Google NLU.