Underspecification in Natural Language Understanding for Dialog Automation

Come join us for our Boston NLP Meet up!

Our special guest is John Chen (https://www.linkedin.com/in/john-chen-76925b1a/) from Interactions

*Pizza and Beer will be provided :)

Underspecification in Natural Language Understanding for Dialog Automation

Chatbots have come to be used for many diverse applications from
responding to information seeking requests to controlling home
appliances. Many chatbots rely on an intent classifier, which they
use in order to understand the user's utterance by classifying it into
one of a finite set of predefined categories. Accurate intent
classification is crucial, since misclassification would adversely
affect user experience. In this work, we explore the strategy of
underspecifying the intent classifier's output when the classifier's
confidence score goes below a certain threshold. When applied to a
SVM intent classifier, we demonstrate that this strategy significantly
reduces the number of misclassifications. In addition, we discuss its
integration into an enterprise-grade chat-based customer care routing
application.

Bio
John Chen is a computer scientist at Interactions LLC, a company that
provides automated customer care services, headquartered in Franklin,
Massachusetts. At Interactions, he has been working on improving the
automation of chat customer care by applying natural language
processing and machine learning techniques. Previously, at AT&T
Labs--Research, he was involved in various projects including building
a real-time incremental speech translation system. At Janya, Inc., a
text analytics company, he worked on various information extraction
problems such as semi-supervised learning of IE modules, and entity
and event coreference. At various other places, he has worked on
domain adaptation for parsing, semantic role parsing, and
summarization of email threads. He earned a PhD from the University
of Delaware, where his thesis work concerned automatic extraction of
tree adjoining grammars from treebanks and their application to
statistical parsing.

Directions to Rakuten: