© August 2018 | IJIRT | Volume 5 Issue 3 | ISSN: 2349-6002
IJIRT 147009 INTERNATIONAL JO URNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 54
Neural Intent Recognition for Question-Answering
System
Ajinkya Pradeep Indulkar
1
, Srivatsan Varadharajan
2
, Krishnamurthy Nayak
3
1, 3
Department of E&C, Manipal Institute of Technology, Manipal
2
Philips Lighting India Limited, Bangalore
Abstract- Conversational Agents, commonly known as
Chatbots, are a successful result of the collaboration of
Natural Language Processing (NLP) and Deep
Learning. Major Technology Giants like Google,
Amazon, Microsoft, etc. are heavily invested in
developing a sophisticated conversational agent which
can pass the Turing Test. There are various
methodologies which can be used to develop the various
components of such an agent from scratch. This paper
uses SQuAD, an open Question-Answering dataset, for
developing an Intent Recognition System for any
Question-Answering system. Inspired by Author-Topic
Modelling, a Title-Topic Modelling technique is used in
combination with various Deep Learning models to
train the Intent Recognition System, achieving an
accuracy of 88.36%.
Index Terms - Conversational Agents, Deep Learning,
Intent Recognition, Natural Language Processing,
Question-Answering System, Title-Topic Modelling.
1. INTRODUCT ION
The origin of conversational agents can be traced
back to 1950 when Alan Turing, father of Computer
Science, published his first paper [1]. His question,
“Can Machines think?” ushered the world into the era
of Artificial Intelligence (AI) which began a plethora
of research to turn his question into reality.
Numerous sub-fields of AI like Machine Learning,
Computer Vision, and recently, Deep Learning have
seen more researchers committing themselves to
develop novel techniques in the respective fields.
The development of conversational agents falls under
the field of Natural Language Processing, which is an
area of research concerned with the understanding of
human (natural) languages by the machines. Similar
to Artificial Intelligence, its origin can also be traced
back to 1950s when Alan Turing proposed a criterion
for intelligence, known as the “Turing Test ”.
A machine passes the test when a human is unable to
distinguish the machine‟s responses from that of a
human‟s.
ELIZA [2] was the first implementation of a human-
like dialog system, developed in 1960s. It worked on
a rule-based methodology. Since then, the number of
research papers published in this field is
overwhelming.
NLP Techniques like Dependency Parsing, Named
Entity Recognition, etc. are vital nowadays in the
development of the most sophisticated conversational
agent.
A. Motivation
The current scenario of the chatbot industry mainly
involves detecting the intent of a user utterance to the
conversational agent by extracting entities and using
classic machine learning algorithms such as Support
Vector Machines to develop a classifier for the pre-
defined intents. This is big step from the previous,
rule-based approach but still lacks the true AI
capabilities. Natural Languages are complex, and a
human can express an intent in more than one way.
To be able to understand such complexities can allow
any machine to pass the “Turing Test”.
This paper explores methodologies involved in
developing a Question-Answering system using the
concepts from NLP and Deep Learning.
The Neural Natural Language Understanding engine
uses Intent Recognition System to understand the
intents expressed by the user. For training the Intent
Recognition System, Topic Modelling techniques are
used to automatically label the training data, as the
original dataset is unlabeled.
The techniques used are modular, i.e. they can be
used individually in other systems according to the
researcher‟s requirement.
II. LITERATURE REVIEW