Survey on Chatbot Classification and Technologies

2,3,4,5

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 09 Issue: 11 | Nov 2022 www.irjet.net p-ISSN:2395-0072

Survey on Chatbot Classification and Technologies

Adityan 5

***

Abstract - Wanting to reach out to as many people as possible has always been a primary goal of mass production. Virtual agent chat-bots communicate with many people individually. Chat-bots can answer business questions, help make orders, teach a language or play music. Online businesses have implemented simple answering chat-bots that answer commonly asked questionsandconnectarealhumancustomerserviceagent when a specific or complicated question is asked. Online businesses have also implemented high level NLP systems that automatemany configuration tasks after processing a sentence in natural language. In this survey paper we will see chatbot evolution and their different type of technologies used and compare them to understand the techniques, current best chatbots and their limitations to give an idea for someone to improve certain areas in a chatbottomakethemmorehumanlike.

Words: Chatbots, NLP, Deep Learning, RNN, CNN

1. INTRODUCTION

The idea of a chatbot comes from the imitation game or the Turing test which was created by Alan Turing in 1950. This game aimed to imitate human behavior. In 1966 the first chatbotcalled ELIZA was developed. This system used keyword matching and minimal context identification.

This bot lacked the ability to maintain realistic human conversations. In the 1980s, the ALICE or Artificial Linguistic InternetComputerEntity chatbot wascreated. This bot was considered to be significant due to the use of the Artificial Intelligence Markup Language AIML. The idea behind AIML was to declare the pattern-matching rules which connect user-submitted words and phrases. The Jabberwack chatbot was built to simulate natural human language to learn from previous conversations and then the contextual patterns were used to select the most relevant response. Additionally, commercial chatbots called Lingubots were developed to customize the template to analyze the word structure and grammaroftheuser’sinput.

1.1. Uses of Chatbots

Recently, the importance of chatbots in thepublic sector has taken place. For example, chatbot was used for political purposes to inspire public opinion and intervene in any discussion in social media about politics. Another chatbot has been proposed as a digital channel of communicationbetweencitizensandthegovernment.

Intheeducationsector,achatbothasbeenused to enhance critical thinking and supportlearnersinlearning a new language as the user can learn from the chatbot throughtheirconversations. Aneducational botcombining an intelligent tutoring system and learner modeling was designed to support learners. Another chatbot was proposedformedicalstudentsfor educationalpurposes.In the health care sector, Your.MD chatbot was developed to providerelevanthealth informationforpatients.

Shawar and Atwell developed an algorithm for retraining a chatbot in a specific domain about a specific topic inany language. Their algorithm was appliedontwo different languages, Arabic and Afrikaans, using the different corpus, the Quran to compute frequently asked questionsandthecorpusofSpokenAfrikaans,respectively. Inthepastfewyears,chatbotshavebeenincreasinglyused by several organizations to increase the response time to customers in answering their questions and also reduce operationalcosts.

Chatbot applications have been used in both the private sector, including the virtual assistants that are poweredbyvoice(e.g.Siri,Alexa,Googlenow,Cortana)and public sector gaming agencies, telecommunications, banking (implementing transactions), tourism (booking hotel and travel tickets), media (news provision), retail, stock market and insurance companies. Additionally, governmentshaveusedchatbotsonsocialmediaplatforms suchasTwitterasanewformofpoliticalcommunication.

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 09 Issue: 11 | Nov 2022 www.irjet.net p-ISSN:2395-0072

1.2 Simple Working of Chatbots

We can broadly divide the working of chatbots into 4 stages.

1. Input from a user: Input can be in any form, basically in text, but if input is invoice, then it need to convert into text. Basically a series of wordsasaqueryorjustanormalphrase.

2. Analyze Users query: Chatbot with respectto the technology involved in it tries to analyze and interprettheusersphrases.

3. Identify intent and entities: After analyzing the sentence, it breaks it down the phrase into words and use its knowledge base to identify keywords andintentstomatchthequery.

4. Compose reply: After identifying the entities, it needstocomposeareplythatis appropriatetothe usersqueryandgiveasoutput.

2. CHATBOT CLASSIFICATION ANDTYPES

There has been a number of classifications of chatbots. We can classify the chatbots based on various factors like the technology used in them, the various domains they are being used, or based on their architecture. We can classify chatbots based on technology like AI chatbots, ML based deep learning chatbots, use of NLP or NER in chatbots, etc. We can classifychatbotsbasedontheirfunctionalityforacertain domain like Hospital managing chatbots, Shopping orientated Chatbots, customer service chatbots, etc. We can classify them based on architecture used like a complexCNNorarulebasedone.

Wecanmainlyclassifythechatbotsbasedonthe technology used in them or the way the processing of data happens in them. A simple way of classification on chatbotsintomainly2types,taskorientedandnon-task oriented. Again task oriented can be classified into 2 types, Supervised approach and Unsupervised approach. Further there are two types in non-task oriented, Retrieval-based chatbots and Generation- based chatbots[2].

There is another broad classification of these Conversational Agents. This classification is more like based on their architecture. We can classifythem mainly into 4 types, 1.Interact Mode, 2.Knowledge Domain, 3.based on Goals, 4.Design Approach [3].Interact Mode chatbots are the ones which we use on daily based like Apples “siri”. They can be Text-Based or Voice-Based. There Architecture mainly is complex AI. They continuously develop themselves by machine learning algorithmsbyobservinguserinteraction.Thereisalotof

research is going on them to improve the chatbot responses.

Knowledge Domain based chatbots are more like mainlyfocusesonretrievingappropriateresponsesfrom a databaseona particulardomainorareawhereithassome knowledgeabout bysearching andmatching thekeywords in users questions. These can be Open Domain or Closed Domain. Open Domain is chatbots are the oneswhich have no restrictions on area of expertise or domain knowledge. Theyarenotrestrictedtoonlyone typeofdomain. Instead they can be know about a lot of domains. Due to this feature, the conversation with these bots is more realistic and natural. Where as closed domain chatbots knowledge is limited to a certain extent. They can only answer to the questionsiftheybelongtotheirdomainknowledge

The chatbots based on their Design Approach are of three types, Rule-Based, Retrieval based and Generative Based. Rule Based chatbots responses are all pre-defined, so the developers can control the chatbot conversation levels[7]. It used a tree like structure to answer the user queries with multiple follow-up questions to match and give best response possible. Retrieval based chatbots are like similar to Rule-based with closed domain knowledge and use of neural networks making them advanced chatbots. They mainly do three steps, intent classification, entity recognition, and response selection. They can be implemented with techniques like multi layer perceptron or sentence similarity, etc[8]. Generative based chatbots are advanced bots which can generate response by combinations of language rather than just selecting predefinedresponses.Thesearebuildbyseq2seqmodelsused for machine translation. Deep learningtechniques can be usedtorefinethesechatbots[9].

Now a days, AI and natural language processing(NLP)isbeingusedtocreateadvanced chatbots that talk like a real human. Googles “ok google” and Amazons “alexa” are present advancedchatbots. But they alsohavesomelimitationslikeunderstandingtheintentor users tone and mood in conversations. Yes they do a fine work in giving responses and setting up tasks like alarm butthey might notgoodatunderstandinghumantoneand give suggestions at certain moments. Chatbots with NLP technology is used to analyse a text and interpret it to improve their ability to give responses[6]. They can recognize the sentence structure and determine what the useris saying instead ofjustgivingsome pre-programmed response. These are often referred as AI virtual agents or assistants. They can replace human agentsover repetitive andtime consumingcommunication.

From above, we can now say that they are mainly two types of chatbots, one with intelligence i.e., use of AI and NLP technology in chatbots and another with preprogrammed responses. NLP based are more realistic and more human like because of its deep learning abilities to

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 09 Issue: 11 | Nov 2022 www.irjet.net p-ISSN:2395-0072

give responses in a conversation. Pre-programmed responses chatbots are more like a structured tree like floworinprogrammingtermsnestedif-elseflowcontrol togivebestmatchedresponse

3. AI TECHNIQUES USED IN CHATBOTS

Todesignachatbot,developer mustbeawareof techniques like parsing, pattern matching, AIML, chat script, SQL and relational database, Markov chain and language tricks [5]. Out of these, the AIML is the one whichgiveschatbotsomeintelligencemakingthemmore humanlike.WithTimemanytechniquesofAIandMLare being applied to chatbots to make them more advance andtomakehumanlikeconversations.NLP,NLUandNER are the main technologies that helps chatbots to analyse the text or user queries. While NLP used to interpret the texts,NERisusedtoidentifyandclassifythekeywordsin usersword.

NER (Named Entity Recognition) system processes the sentence to identify and classify the keywords into pre-defined categories or entities. These Entities can be anything like dates, names, locations, Employee IDs, etc. The main component in a chatbot is Natural Language Understanding(NLU) unit. But they take a lot of time to build them from scratch. So integrating various NER modulesintoknowledge base of chatbots can help to build NLU units faster and more efficient[4].

Augmentation capacities in chatbots make them give more than one appropriate response. We need to augment them with certain predictive analysis features, givingthemtheabilitytounderstandusersmood.Itgives ability to ask users when they don’t understand a question and to give another response when a user does not understand aresponse by rephrasing an answer. It also enables them some problem solving ability in chatbots.

Deeplearningrecurrentnetworksmakechatbots more advance. Seq2Seq is a advanced model of chatbots also known as Encoder-Decoder Model. It contains encoder and decoder for machine translation. It uses recurrent neural networks to process series of text or words in a sentence and trains the model. It is the best model for conversational chatbots. It not only takes currentinputintoaccountbutalsotheneighboringwords also to generate responses by taking sequence of words as input. It generates a series of words as its output which it uses again as an inputin next step. This takes twoinputs ata time, oneis from itspreviousproduction andotherisfromtheuser[1].

4. EFFECT OF TECHNOLOGIES ON CHATBOT PERFORMANCE

Therehasbeenanumberofadvancedtechnologies came into use in chatbots to make the advanced. From Eliza, which worked on words matching over pre-defined responses to current Alexa, which is advanced by use of deep learning and NLTK techniques, there has been a numberofmethodsinvolved.

4.1 Words Matching Over Scripted Responses

Eliza iscreatedusingthistechnique.Itis naturally pre-programmedhardcodedresponsesinthedatabaseand themachinewillmatchthe words ofuserinputoverthem. Butithassomedrawbackslike,identificationofkeywords, transformationofresponses,discoveryofminimal context, generationofresponsesfornonmatchedkeywords[10].

4.2 Artificial Intelligence MarkupLanguage(AIML)

Richard Wallace developed AIML in 1995which is the forming foundation for A.L.I.C.E chatbot. The techniques used in developing various A.L.I.C.E prototypes didn’t involved complex machine learning techniques and sophisticated natural language processing. A software is developed to convert readable text(corpus) into AIML format. It works by generating AIML knowledge Base Automatically. The limitation for this is lack of manual developmentof its knowledge[11].

4.3. Deep Learning Methods

Deeplearningismainlyaboutgivingamachineanartificial brain which consists of artificial neural networks to process the data like human brains. The methods like multi-layerperceptron,sentencematchingandseq2seqare models to develop chatbots using deep learning. The techniqueslikeBeamSearchDecodingmakesthesearching of answers for questions easy thus increasing the performance of chatbot. It uses abreadth-first Search to build its state-space tree. It works in a greedy approach [12].Limitationsforthistypeofchatbotsaredependsupon the algorithm used to develop them. But some common limitations like not recognizing voice correctly might be a huge problem to solve to increase the efficiencyand usage ofchatbot.

4.4.Effect of NLP on Performance of Chatbots

Natural Language Processing(NLP) gives machine theabilitytoidentifypatternsof wordsandinterpretthem appropriately, giving the chatbot the ability to have conversations with humans more effectively. It is like an extra layer on top of the deep learning recurrent model architecture. But limitations like Grammatical Errors, Semanticsandaccuracyhavetobeimprovedbytuningthe model more. Use of techniques like Feedback Mechanism

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 09 Issue: 11 | Nov 2022 www.irjet.net p-ISSN:2395-0072

and Policy Learning in Agent for Dialogue Management can manage the real context of the user saying. The Questionandanswer Systemmakes a chatbot toidentify frequently askedquestionsandgiverepliesimmediately. ItcanbedonebyManualTrainingorAutomatedTraining [13].

5. PERFORMANCE METRICS FORCHATBOT

Wecananalyzeandmeasuretheperformanceof a chatbot by various metrics. Some basic metrics are Speed,Interoperability,turingtestandscalability.Wecan also use traditional R- squared and RMSE factors to evaluate responses. Some Key performance indicators are:

1. Bounce rate: volume of sessions wherethe chatbot wasopenedbutnotused

2. Satisfaction rate: average grade given when evaluatingthechatbot’sanswers(tobalanceoutwith theevaluationrate).

3. Evaluation rate: percentage of user sessions that have given an evaluation of the chatbot’s answers at leastonce.

4. Average chat time: allows you to evaluate your users’interestforyourchatbot.

5. Average number of interactions: used toevaluate theCustomerEffortScoreonthechatbotandmustbe correlated to the satisfaction rate. If the latter is very low, the bot may be engaging the users in too many branchesandstepstomeettheirneeds.Inthiscase,a resolution can be to correct the decision trees or knowledgebasearchitecture.

6. Goal completion rate: in case your bot contains targeted actions like CTAs, a form or some crossselling,thatistherateofuserswhohavereachedthat specificactionthroughthechatbot.

7. Non-response rate: the amount of times the chatbot has failed to push some content following a user question (due to lack of content or misunderstanding).

Use of these KPI’s doesn’t measure performance of a chatbot completely, but they are sufficient to improve at various limitations of chatbots. Further metrics can be vary depending upon on which domain the chatbot is being used like health-care, shopping or customer service.For domainspecific metricsweneed toconsider metricslike:

Health-care: correct diagnosis, correct prescriptionsfor diseases,etc

Shopping: level of accurate suggestions and conversion rateofsales,etc.

Customer Service: Level of answering customers queries andaccessibility,etc.

6.CHATBOTEVOLUTION

To test a chatbot, turing test is used. If it passes then chatbotiscableofhavingarealhumanconversation.

In 1966 first chatbot has been introduced into the world, created by Joseph Weizenbaum called ELIZA. It uses patternmatchingandsubstitutionmethodology.

PARRY another chatbot built in 1972 by a psychiatrist. Thisprogramisusedtosimulateadisease.

JABBERWACKY wascreatedbyRollo Carpenterin1988.It usedanAItechniquecalledcontextualpatternmatching.In 1992 a full voice operated chatbot was developed called Dr.Sbaitso bycreativelabs.

A.L.I.C.E(ArtificialLinguisticInternetComputer Entity)isa universal language processing chatbot that uses heuristic pattern matching to carry conversations developed by RichardWallacein1995.

SIRI developed by Apple as voice assistant for iOSusers in 2010. It uses natural language UI and also can perform varioustasks.Commandtoactivateis“HeySiri”.

GOOGLE ASSISTANT waslaunchedin2012 by Googleit is similar to Siri but for android users. It can make calls searchtheinternetforanswers.Commandtoactivateis“Ok Google”.

CORTANA introducedbyMicrosoftin2014and integrated with windows 10. It uses voice recognition and various algorithms to give answers.Search box can also be used to chatwithCortana.Commandtoactivateis“HeyCortana”.

ALEXA is a personal assistant developed by Amazon and introduced in 2014. It is built into devices like Amazon Echo and Echo Dot. Command to activate is “Alexa” followedbyanycommand.

7.CHATBOT COMPARISIONS

To check chatbot performance we have taken 4popular chatbots to check their performance. SIRI, GOOGLE ASSISTANT, CORTANA, ALEXA. These chabots are popularly used all over the world. These chatbot are subjected to different commands and their response is recorded. These chatbots learn overtime based on their usagetogiverelevantanswerstotheuser.TestResultsmay waryfromusertouser.

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 09 Issue: 11 | Nov 2022 www.irjet.net p-ISSN:2395-0072

Commands:

1. Singasongforme,willyou? 2. Couldyousuggestmeanygoodrestaurant. 3. Whatareyourhobbies? 4. Iwouldliketodrinkwine.Whatdoesittastelike? 5. Canyoufillmyresumeusingagoodtemplate? 6. Switchoflightsformeinthelivingroom.

Table-1: Resultsofqueries

Ok Google Siri Cortana Alexa Q. 1 Customlyrics everydaywith goodtune.

Q. 2 Findsby looking throughthe browser

Q. 3 Itisable answer withsome hobbies

Customlyrics buthasno tune.

Isableto lookupnearby restaurants ratingand Directions

Lyricsfrom songfound online.

Isablelookup restaurants throughbing

Singing randomsong

Suggested Restaurant.

Cannot answer Itisableto answer Responded with hobbies

8. CHATBOT APPLICATIONS

8.1.

Health Care Domain

Before you go to the doctor, majority of us google the symptoms .In fact 89% of patients google their health symptoms before scheduling an appointment this isn't goodideaandresultsarebad.

Luckily chatbots realize the problem by opening up a app, diagnose yourself by chatting with AI driven health companion

1. Babylon-Itisdesignedaroundadoctor'sbrain,using ArtificialIntelligence(AI).

2. Elomia-is an AI powered therapist. It understand, listen and support. The company trained the AI through observationsofpasttherapies

3. Yuper- it offers personality assessments, moodtracking, emotional tests,8 of 10 users reported an improvedmoodafterjustoneconversation.

8.2.Customer Service

Q. 4 Searchesfor keywordslike wineandtaste

Searchesfor keywordslike wineandtaste

Searchesfor keywordslike wineandtaste

Suggested different typesof winesand theirtastes Q. 5 Suggestsa template frominternet

We live in a world of instant messaging and communication, where customers want instant help and support. If you don't provide quick and appropriate customer support you lose customer loyalty and if you try to do everything with customer service they end up answering the same questions and wasting money on agents.Sodeployingachatbotismoreefficient.

Suggestsa template frominternet

Unableto answer thequestion

Suggested webpages Q. 6 Canperform IoT tasks

1. Zoho desk-Best for service teams looking for a tool thatlearnsfastfromtheknowledgebase

Canperform IoT Tasks

Cannot performIoT tasks

Table-2: ChatbotPerformanceTable

Switchedoff the lights. Ok Google Siri Cortana Alexa Relevance(6) 3 3 15 4

2. Intercom-Bestforteamsthatwanttocreatetargeted messagesfordifferentcustomersegments

3. IBM Watson Assistant-Best for customer service teamsthatwanttoreducecosts

8.3. Personal Assistants

Apersonrequiresapersonalassistantinhis/herbusylife to remind or follow their work routines to achieve their dailygoals. 1.Siri 2.Alexa

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 09 Issue: 11 | Nov 2022 www.irjet.net p-ISSN:2395-0072

9. CONCLUSION

We presented the past works, evolution of chatbots, classification of them and the technologies emergenceandtheireffectsonchatbotefficiency.Theirhas beenanumberofimprovementstomakethechatbotmore human like. At present Deep learning Recurrent Models are leading and dominating the technology used in chatbots. We can classify the chatbots mainly into two types,onewithintelligencei.e.,whichcanprocessthedata and generates its own responses, another is with no intelligencei.e.,pre-definedknowledgebasemodels.Notall chatbots require that advanced technology. It depends on usage of chatbot, we can choose from pre-defined responses model to recurrent models. The metrics used to measure the efficiency also depends on the domain the chatbot is being used. From the comparisons of chatbots, we can conclude that every chatbot has their own limitationsandALEXA,OkGoogleare leadingchatbots. The success of those chatbots is may be because of their analysisofuserinteractionandcontinueslearningprocess.

10. FUTURE WORKS AND IMPROVEMENTS

One can really try to tune a chatbot more to enhanceitsabilitytohaveaconversationwithhumans.We can develop a human mimicking chatbot that can chat on behalf of a person with different people and different tone in theconversations depending upon the relationship of those peoples. It can be made my analysis of user conversationswithdifferentpeopletoidentifypatternsand level of tone with different people. If we can achieve a chatbot that can actually mimic a human, then we can use that technology tointegrateitwithhumanoidrobotsalso.

ACKNOWLEDGMENT

We would like to thank our guide Dr. M. V. VIJAYA SARADHI and project coordinator Mrs. Soppari Kavitha for guidance, without their guidance, the research on survey would not have been completed successfully. We are extremely grateful to Dr. M. V. VIJAYA SARADHI, Head of the Department of Computer Science and Engineering. Ace Engineering College for his valuable and constant support throughout the execution of this work.

REFERENCES

[1] Satyendra Praneel Reddy Karri, Dr.B.Santhosh Kumar: Deep Learning Techniques for Implementation of Chatbots, Conference: 2020International Conference on ComputerCommunicationandInformatics(ICCCI).

[2] Ebtesam H. Almansor, Farookh Khadeer Hussain.: Survey on Intelligent Chatbots: State-of-the-Art and Future Research Directions, In book: Complex, Intelligent,andSoftwareIntensiveSystems(pp.534-543).

[3] Omid Ameri Sianaki, Nedal Ababneh.: A Survey on Conversational Agents/Chatbots Classification andDesign Techniques. In book: Primate Life Histories, Sex Roles, andAdaptability(pp.946-956).

[4] Nazakat Ali.: Chatbot: A Conversational Agent employed with Named Entity Recognition Modelusing Aritificial Neural Network., source: arxiv.orgcs.

[5] Sameera A. Abdul-kader, Dr. john Woods.: Survey on Chatbot Design Techniques in Speech Conversation Systems, (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 6, No. 7,2015.

[6] NiranjanDandekar,SuyogGhodey.:implementationof chatbot using Natural Language Processing,International Journal of Advance Research in Science and Engineering, volno.6,issueno.12,Dec2017.

[7] Fikadu Wayesa.: Design and Implementation of a Rule Based Afaan Oromoo Conversational Chatbots, International Journal of Scientific Research in Science and Technology, Vol 5, Issue 8, NovemberDecember-2020.

[8] Andrew Ciayandi, Viny Christanti Mawardi, Janson Hendryli.: Retrieval Based Chatbot on Tarumanagara University with Multilayer Perceptron, IOP Conference Series,2020.

[9] Salim Akhtar Sheikh, Vineeta Tiwari, Sunita Bansal.: GenerativemodelchatbotforHumanResourceusingDeep Learning,ICDSE,2020.

[10] Joseph Weizenbacm.: Eliza- A Computer Program For the Study of Natural Language Communication Between Man and Machine, Vol 9,Issue1,Jan-1996.

[11] Bayan ABuShawar, Eric Atwell.: ALICE Chatbot: Trials and Outputs,Computacion y Sistemas, Vol 9, No 4(2015),ISSN2007-9737.

[12] Satyendra Praneel Reddy Karri ,Dr Santhosh Kumar.: Deep Learning Techniques for Implementation ofChatbots,ICCCI-2020.

[13] Soufyane Ayanouz, Boudhir Anouar Abdelhakim, Mohammed Benhmed.: A Smart Chatbot Architecture based NLP and Machine Learning for health care assistance,NISS-2020.

[14] Shivang Verma, Lakshay Sahni, MoolchandSharma.: Comparative Analysis of Chatbots, IRJET,Vol7,Issue12, Dec2020.

[15] Vijayaraghavan V., Jack Brian Cooper, Rian Leevinson J.:Algorithm Inspection for Chatbot PerformanceEvaluation,CoCoNet’19.