ijeter35852020.pdf

Kolla Bhanu Prakash et al., International Journal of Emerging Trends in Engineering Research, 8(5), May 2020, 1709 – 1715
1709

ABSTRACT
The goal of this project is to develop a chatbot using deep
learning models. Chatterbot is an existing research area
whose main goal is to appear as human as possible and most
of the current models(which use RNN and related sequential
learning models) are unable to achieve this task to relate over
long dependencies. Adding onto that , NLP tasks require a lot
of data which can be hard to collect for smaller projects/tasks
which inspired to tryout sequence to sequence learning model
using LSTM. For that we have used a movie dialog corpus of
220,579 conversation exchanges among which about 50,000
conversational exchanges are onlyused as a training corpus to
our model since training on more conversation exchanges
requires high computation power than we have.
Key words : Deep learning, Chatterbot, RNN, NLP, LSTM.
1. INTRODUCTION
A chatbot is a software program that uses text or
text-to-speech to perform an online chat conversation instead
of direct communication with a live human user. Built to
convincingly mimic how a human being will haves
conversation with partner, chatbot systems usuallytesting and
tuning in continuous way, and many in development remain
unsuccessful effectively pass after conversing the standard
Turing test . In 1994, Michael Mauldin (creator of the first
Verbot) originally coined the word "ChatterBot" to describe
such conversational programs.
Figure 1: Chatting Interface (Chatbot Interacts)
The Figure 1 shows the chatting interface for the user and
chatbot interaction and the Figure 2 shows platforms/apps
that uses chatbot. Natural language processing is one
important area in AI research. AI fields that are weak usually
implement advanced applications or programming languages
designed specific to the limited role they require. For
example, markup language used by A.L.I.C.E. (AIML) that is
unique to its use as a chatting agent and has been used by
Alicebots. Nevertheless, A.L.I.C.E. is still is focused on
pattern matching techniques and no practical experience, as
back in 1966 ELIZA used the same methodology [12].
Figure 2: Platforms/Apps that uses chatbot
Integrating NLP in the chatbots means more human presence.
If a chatbot is created and deployed , this is for general usage,
so you also see people asking questions about it. It seems very
much in line with human nature that users would try to fool
and throw off the chatbot. You may attempt to solve that by
adding default responses, but that tends to fall short very
much as it is almost difficult to anticipate which questions
will be answered and how they will be raised as shown in
Figure 3. Latest research focused mainly on unsupervised and
semi-supervised algorithms in computing.
Figure 3: NLP for Chatbots
Chatterbot implementation using Transfer Learning and
LSTM Encoder-Decoder Architecture
Kolla Bhanu Prakash1
, YVR Nagapawan 2
, N Lakshmi Kalyani 3
, V Pradeep Kumar 4
1
Assoc.Prof.- Department of CSE, Koneru Lakshmaiah Education Foundation , India, drkbp1981@gmail.com
2
Assi.Prof.- Department of CSE, Anurag Engineering College, India, ynpawan@gmail.com
3
Assi.Prof.- Department of CSE,VNR Vignana Jyothi Institute of Engineering & Technology, India,
neerukondait@gmail.com
4
Assi.Prof.- Department of CSE, B. V. Raju Institute of Technology, India, pradeepvadla@gmail.com
ISSN 2347 - 3983
Volume 8. No. 5, May 2020
International Journal of Emerging Trends in Engineering Research
Available Online at http://www.warse.org/IJETER/static/pdf/file/ijeter35852020.pdf
https://doi.org/10.30534/ijeter/2020/35852020

1710
Today's AI systems are capable of interacting with
consumers, assessing their desires, identifying their
preferences and recommending an acceptable course of action
with minimal or no intervention from person. Many
conversational agents accessible now, such as Apple's Siri,
Microsoft's Cortana, Amazon's Search Assistant and Alexa.
The basic principle of chatbots is to have the best answer to
every question it receives. [5]. The Figure 4 and 5 below
shows Chatbot's conceptual map using Deep learning.
Figure 4: Chatbot conceptual map
Figure 5: Classification using Neural Network
Nearly all chatbots use deep learning, some version of a
Sequence to Sequence (Seq2Seq) pattern. In 2014, with a
paper named "Sequence to Sequence Learning with Neural
Networks," Ilya Sutskever, Oriol Vinyals and Quoc Le
published the foundational work in this area.
RNNs are a neural networks that can manage sequential data,
such as videos (frame series), and more commonly, sequences
of texts or virtuallycollection of symbols. The benefit of this is
that the network doesn't need to ask what the icons say. Think
of the condition as network memory as in this memory unit
(internal state), which is constantly modified during each
process, RNN devours a sequence (sentence), word by word,
important information about the sentence. The main
distinction between a normal RNN network and LSTM / GRU
networks is the architecture having the memory unit Figure 6.
An LSTM cell has multiple gates to keep track of useful
information, to forget unnecessary information and to
carefully expose information at every move.
Figure 6: LSTM Architecture
The Figure 7 Sequence to Sequence model(seq2seq) is made
up of Two RNN encoders, and one decoder. The encoder
reads the list of inputs word for word and transmits a
description, ideally catching the input sequence's context.
The decoder generates the output sequence. The goal is to
jointly optimize the log probability of the output series
according to the input series [6].
Figure 7: Sequence to Sequence Model
Cho and. Al. suggested a novel form of the neural network
named RNN Encoder-Decoder consisting of two recurrent
neural networks (RNN). One RNN codes a symbol sequence
into a representation of a fixed-length matrix, and the other
decodes the representation into another symbol sequence.
1.1 Transfer Learning in Chatbot
In training deep neural networks, AI engineers have been
increasingly excellent at correctly mapping from inputs to
outputs involving broad data sets whether they are words,
images, mark predictions etc. However, the ability to take a
broad perspective of situations that aren't similar to those they
experience during training is often missing in these models.
Real world simulations can be chaotic, and engineers are
forced to make incorrect assumptions if they apply these
models to data sets that have not been observed. The Figure 8
Transfer learning should solve the problem.
Transfer learning is basically a model's capacity to transform
information into new environments. This work implements a
modern Transfer-Transfo method, for conversational
generative data-driven systems such as chatbots. The
algorithm incorporates programming based on transfer.

1711
Figure 8: Transfer Learning Model
learning as well as transformer models Like Figure 9. It
performs fine-tuning for the algorithm by using a multi-task
objective which incorporates a variety of unsupervised
prediction procedures. Supervised learning, powered by deep
learning, is largelyresponsible for the wave of AI applications
that we have seen in the last few years [21]. Transfer learning
will make these developments significantly greater.
Figure 9: Transfer-Transfor input representation
Pretrained language models were used to produce
state-of-the-art outcomes on a wide variety of NLP tasks (e.g.,
sequence marking and grouping of sentences). Any of the
latest works using pre-trained language models include the
transformer ULMFit, ELMo, GLoMo and OpenAI. These
language modeling systems perform extensive pre-training of
hierarchical representations or graph-based representations
of whole structures. This idea represents a change from the
usage of unary word embedding, which has been used
extensively over the past few years to solve many NLP
functions, to the use of more complex, abstract
representations [9].
Figure 10: Evolution of Transfer Learning in NLP
Pre-training allows a model to catch and absorb from a
large-scale corpus a range of linguistic anomalies, such as
long-term dependency and negation [22]. This information is
then used (transferred) to initialize and then train another
model for performing well on a similar NLP task, such as
classifying sentiments as shown in Figure10.
2. RELATED WORK
2.1. A Chatbot Using LSTM-based Multilayer
Embedding for Elderly Care
Because of population changes, programs targeted to elderly
people are now more important than ever, and increasingly
available. The details from the MHMC's daily encounters
with older people have been gathered in this report. Because
individuals may say the method, the sentences that are
recorded in the preprocessing phase to cover the variation of
the conversational sentences are translated into patterns. A
multi-layer embedding technique based on LSTM is then used
to extract contextual details between terms and phrases using
several sentences in a single turn when communicating with
the elderly. Ultimately, the Euclidean distance is used to
choose an appropriate type of question, which is then used to
choose the correct response for the elderly [8].
Figure 11: Chatbot Framework
This research launched a chatbot for elderly treatment, which
can be used to assist Or talk with seniors to relieve their
loneliness. They compiled the collection of chitchat from 40

1712
topics in order to perform a daily discussion as training
material. All the messages were classified and patterned in
the MHMC chitchat database As shown in Figure 11. Then
they used a two-layer LSTM model to address the interaction
between words and phrases in the semantics. Finally they
choose the corresponding response by contrasting pairs of
message through Euclidean distance. The outcome indicates
that a stronger answer can be generated by the suggested
approach, achieving 79.96 per cent matching the top 1
message pair [10].
2.2. A Chatbot by Combining Finite State Machine,
Information Retrieval, and Bot-Initiative Strategy
They developed a conversational framework in this work by
integrating a finite state process, a model of retrieval and a
technique of dialog between computer and initiative. Unless
the user's utterance is beyond the range of modelled
dialogues, the input is interpreted by the retrieval algorithm.
This model of retrieval has demonstrated greater consumer
satisfaction in comparative experiments than the one centred
on the neural-network. Yi et. Al used data from different
sources to create a conversational model. Firstly, they used
Twitter scrapped Evi and dialog details to handle everyday
conversations.[2]. They identified 31 transitions and made 12
states using model of ASK purpose to model dialogues [2]. It
enters 'Start' state when their program is initialized. A
solution is selected between the candidate sentences
depending on the transition triggered by a user request, and
the particular state that stores the background knowledge.
There are two types of dialogs that are handled using finite
state method: multi-round and one-round dialogue [15].
2.3 Learning Phrase Representations using RNN
Encoder–Decoder for Statistical Machine Translation
In this model they proposed a new neural network
architecture named RNN Encoder– Decoder composed of two
recurrent neural networks (RNN). One RNN encodes a series
of symbols into a representation of a matrix of defined length,
and the other decodes the representation into another symbol
set. The performance of a predictive machine translation
system is empirically shown to be increasing utilizing the
conditional probabilities of phrase pairs computed by the
RNN Encoder – Decoder as additional component of the
existing log-linear model [3].
Figure 12: FSM Chatbot Architecture
Cho et Al taught the model to realize the probability of
translating an English word into a matching French
expression as shown in Figure 12. The pattern is then
included in the table of sentences as part of a standard
phrase-based SMT system by rating each pair of phrases. The
analytical assessment shows that this method of rating pairs
of phrases with an RNN encoder – decoder increases the
efficiency of the translation [14]
Nguyen et. Al. devised and evaluated Pytorch on deep
learning framework with the utilization of GPU. The model is
an end to end [7], with no hand-crafted rules. Model is
developed from a limited dataset and also provides user
responses. Nonetheless, answers generated do need to be
strengthened to have a constructive conversation. Bahdanau
et al. (2015) recently effectively implemented such a research
method on communication of neural machine by mutual
learning how to manage and perceive words together. [4]. The
list of Vietnam comprises of 1331 questions and responses.
The conversational datasets with 300,000 are pretrained on
measurements out of 1331 pairs of sentences. It obtained 887
pairs of comments and responses after checking and picking
15-word long phrases (including punctuation). 843 phrases
are used in the vocabularyand 918 phrases are included in the
term response. The training time of the model is 288 minutes
and one second, and Google Colab (free GPU) was used [19].
3. PROPOSED ALGORITHMS
The Retrieval-based models select an answer based on the
question from a set of responses. It produces no new words, so
We don't have to owe grammar a look [20]. The Generative
Models are really wise. They generate a response, word by
word, which is dependent on the query. The answers provided
are vulnerable to grammatical errors because of this. These
models are challenging to practice, because they have to study
the proper sentence form on their own [11]. However, when
trained, the generative models outperform the retrieval-based
models in terms of managing previously unseen queries and
produce the user's intention of talking to a human (maybe a
toddler). The proposed model of the chatbot is done using the
Sequence to Sequence model with transfer learning. Here, the
movie dialog is used as a training corpus [18]. This repository
includes a wide selection of metadata-rich fictitious
interactions taken from the scripts of raw films. It has 220,579

1713
conversational exchanges between 10,292 pairs of movie
characters and includes a total of 304,713 utterances of 9,035
characters from 617 movies while only 50,000 conversational
exchanges were used for preparation. We used pre-trained
word embedding from the language model such as openai gpt
for transition learning, and in sequence to sequence learning
model utilizing LSTM encoder and decoder shown in Figure
13, and the processes in between [25].
Line to Line Paradigm applied with RNN Predictive Machine
Translation Encoder-Decoder (cho et. al.) in Studying Phrase
Representations has since been the Go-To Norm for Dialog
Systems and Computer Translation. It consists of two RNNs,
one Encoder and one Decoder. It is constructed of two RNNs.
The encoder takes a number (sentence) as input and processes
at each stage one symbol (word). The job is to turn a symbol
series into a vector that only covers important details in a
series without missing unnecessary information [3].
Figure 13: Sequence to sequence Model (LSTM encoder-decoder)
3.1 Working of LSTM cell
LSTM networks are extended RNNs consisting of extended
memory cells known as gated cells which allow the inputs to
be remembered for a long time [16]. The information in the
LSTM gated cell memory may be retained or discarded and
can be calculated by the weights / value allocated by the
algorithm given, i.e. it learns the value of the information
over time. The relation between the present layer output and
next layer input shows the cell's recursive existence. This
allows for knowledge to be retained in the LSTM cell from
previous periods. In total there are three gates in LSTM:
 input gate
 forget gate
 output gate
Input gate decides whether to allow in fresh data or not.
Forget gate determines whether to remove unhelpful details.
If the forget gate performance is 1, the knowledge is retained
in the cell state and is indicated to be forgotten by closer to 0.
Input gate chooses to let the forget gate input influence the
input at the current time stage and it defines how much of the
activation of each device is retained [16].
The proposed framework methodology consists of different
steps, such as raw data selection, pre-processing of data,
extraction of features and NN preparation. LSTM consists
activation functions such as
 Sigmoid
 Tanh.
Sigmoid feature wants to change the correct details and only
installs memory but cannot remove / forget memories on its
own [13]. "Tanh" controls the values that pass across the
network, which allows the cell satellite to lose information.
Let it, ot, ft reflect the input – output activations, and forget
gates at time phase t, respectively.
Xt = (x1, x2, x3, ..., xn) (1)
ht is the hidden state
ht = (h1, h2, h3, …., hn) (2)
And ct is memory cell that represents
ct = (c1,c2,c3, ... cn) (3)
The current cell output ht is as follows ht = ot tanh (ct)
 New memory (ct)
 current output (ht) next step of time and it repeats
3.2 Word Embeddings
Term Embedding is a technique to achieve dense term
representation in a small vector space. Increasing term,
represented by a fixed-length vector, is a space point which
can be seen. This technique captures semantic connections
between the words. There are some fascinating properties of
the term vectors. Word Embedding is usually performed in
the network's first layer: Embedding layer, in which a word is
mapped (index to word) from a vocabularyto a dense vector of
the given scale. In the seq2seq model, the embedding layer
weights are trained in accordance with the model's other
parameters [1] shown in Figure 14.
Figure 14: Word Embedding graph

1714
4. RESULTS AND DISCUSSIONS
4.1 LSTM Encoder Decoder Architecture
The results are for random inputs from the test data.
RMSE for training dataset = 6.42
RMSE for testing dataset =18.24
The model has shown to respond well for all answers with
RMSE below 20 for both train and test data while it was above
50 for ANN. Compared to generic answers a normal ANN
gives, the answers given by this encoder decoder structure
using LSTM gave a better dynamic answers, albeit , still
meaningless [24]. The grammar is more accurate compared to
basic generative models when manually evaluated shown in
Figure 15.
While the testing and training loss explains how well a model
learns compared to another model, when the task is about
conversational agents these losses are stripped of their
meaning [17]. The main problems with generative type
conversational agents occur once they are outside the testing
and training phases. These phases only evaluate a model
based on how much its replicating the output. A chatterbot is
good only if it works the same way for messages it has never
seen before. This is extremely hard if a model does not
“think” before it answers shown in Figure 16. In our case, the
LSTM, although it beats other generative models, it still
doesn’t have any mechanism to think [23].
Figure 15. Result Screen 1
Figure 16. Result Screen when input is taken from the user
5. CONCLUSION
LSTM based encoder decoder structure has shown to be more
robust, cleaner and faster than a ANN model. This model also
shows promise as a generative based model which prior
generative approaches were not able to show. The sentence
structure is more dynamic albeit still mostly meaningless as it
is free from context when building sentences. But the biggest
advantage shown by this model is the relative ease of
building. All this model needs to work well is lots of data
unlike older approaches which require constant human
intervention to work properly. This clean approach is
desirable and favourable for wide usages of chatterbots.
REFERENCES
1. Su. Ming, Wu. Chung, Huang Kun, H. Qian and Hsin
Wang, “A Chatbot Using LSTM-based Multi-layer
Embedding for Elderly Care”, IEEE Conference on
Orange Technologies, 2017.
https://doi.org/10.1109/ICOT.2017.8336091
2. Yi Sanghyun and Kyomin Jung, “A Chatbot by
Combining Finite State Machine, Information Retrieval,
and Bot-Initiative Strategy”, Semantic Scholar, 2017.
3. Kyunghyun Cho, B. Dzmitry, B. Fethi, Yoshua Bengio,
“Learning Phrase Representations using RNN Encoder-
Decoder for Statistical Machine Translation”, [online:
https://arxiv.org/abs/1406.1078]
4. Trang Nguyen and Maxim Shcherbakov, “A Neural
Network Based Vietnamese Chatbot”, IEEE Int. Conf.
on System Modeling & Advancement in Research
Trends, 2018.
https://doi.org/10.1109/SYSMART.2018.8746962
5. “Overview of deep learning”, 31st youth academic
annual conference of Chinese association of automation,
Wuhan, China, Nov 2016.
6. Alex Sherstinsky (9 Aug 2018). “Fundamentals of
Recurrent Neural Network and Long Short-Term
Memory”, [Online]. Available: https://arxiv.org/abs /1808.
03314
7. “Chatbot Using Gated End-to-End Memory Networks”,
International Research Journal of Engineering and
Technology (IRJET), 2018, India.
8. Vaswani et. al. “Attention is all you need.”, [online: https:
//arxiv.org/abs/1706.03762]

1715
9. Chuanqi Tan, Fuchun Sun, Tao Kong, Wenchang
Zhang, Chao Yang, Chunfang Liu,“A survey on Deep
Transfer Learning”, IEEE Int. Conf. on Artificial Neural
Networks(ICANN 2018).
10. Bayu Setiaji and Feery Wahyu Wibowo, “Chatbot Using
a Knowledge in Database: Human-to-Machine
Conversation Modeling”, IEEE Int. Conf. Intelligent
Systems, Modelling and Simulation, 2016.
https://doi.org/10.1109/ISMS.2016.53
11. AM Rahman, Abdullah Al Mamun, Alma
Islam,“Programming challenges of chatbot: Current and
future prospective”, IEEE Conf. on Humanitarian
Technology, 2017.
12. Hiremath et. al., “Chatbot for education system”,
International Journal of Advance Research, Ideas And
Innovations In Technology, vol 4, Issue 3
13. Kolla, B.P., Raman, A.R.” Data Engineered Content
Extraction Studies for Indian Web Pages”, Advances in
Intelligent Systems and Computing, vol 711, 2019.
https://doi.org/10.1007/978-981-10-8055-5_45
14. Prakash, K.B., Rangaswamy, M.A.D., Raman, A.R.
“Statistical interpretation for mining hybrid regional web
documents”, Communications in Computer and
Information Science, vol 292 CCIS, 2012.
https://doi.org/10.1007/978-3-642-31686-9_58
15. Prakash, K.B., Rajaraman, A.” Mining of Bilingual
Indian Web Documents”, Procedia Computer Science,
vol 89, 2016
16. Ismail, M., Prakash, K.B., Rao, M.N, “Collaborative
filtering-based recommendation of online social voting”,
International Journal of Engineering and
Technology(UAE), vol 7, issue 3, 2018.
https://doi.org/10.14419/ijet.v7i3.11630
17. Prakash, K.B, “Content extraction studies using total
distance algorithm”, Proceedings of the 2016 2nd
International Conference on Applied and Theoretical
Computing and Communication Technology, iCATccT
2016, 2017
18. Prakash, K.B., Rajaraman, A., Lakshmi, M,
“Complexities in developing multilingual on-line
courses in the Indian context”, Proceedings of the 2017
International Conference On Big Data Analytics and
Computational Intelligence, ICBDACI 2017, 2017
19. Prakash, K.B., Kumar, K.S., Rao, S.U.M, “Content
extraction issues in online web education”, Proceedings
of the 2016 2nd International Conference on Applied and
Theoretical Computing and Communication
Technology, iCATccT 2016, 2017
https://doi.org/10.1109/ICATCCT.2016.7912086
20. Prakash, K.B., Rajaraman, A., Perumal, T., Kolla, P.
“Foundations to frontiers of big data analytics” ,
Proceedings of the 2016 2nd International Conference on
Contemporary Computing and Informatics, IC3I 2016,
2016
21. Prakash, K.B., Ananthan, T.V., Rajavarman, V.N,
“Neural network framework for multilingual web
documents”, Proceedings of 2014 International
Conference on Contemporary Computing and
Informatics, IC3I 2014, 2014
https://doi.org/10.1109/IC3I.2014.7019797
22. Prakash, K.B., Rangaswamy, M.A.D., Raman, A.R,
Attribute based content mining for regional web
documents, IET Seminar Digest, vol 2013, issue 8, 2013
23. Ratnam, D.V., Gutta, R.C., Kumar, M.R., (...), Pavan,
P.S., Santhosh, C., ” Effect of high-k dielectric materials
on mobility of electrons”, International Journal of
Emerging Trends in Engineering Research8(2), 2020.
24. Suresh, B., Nikhilesh, T., Abhishek, T., (...), Chandra
Sekhar Yadav, G.V.P., Ghali, V.S, “Qualitative
subsurface analysis in quadratic frequency modulated
thermal wave imaging”, International Journal of
Emerging Trends in Engineering Research, 8(1), 2020.
25. Suresh, B., Sathvik, K.N., Imran, S., (...), Vijayalakshmi,
A., Ghali, V.S, “Diagnosing osteoporosis through
numerical simulation of bone sample by non-stationary
thermal wave imaging”, International Journal of
Emerging Trends in Engineering Research, 8(3), 20202.

ijeter35852020.pdf

More Related Content

Similar to ijeter35852020.pdf (20)

Recently uploaded (20)

ijeter35852020.pdf