ABSTRACTNatural language processing (NLP) is a recent developingarea of research and innovation which is        gaining a lot of popularity these days. Natural language processing is apart of AI which focuses on human like natural language processing and isextensively used in developing virtual assistants and chat bots .Severalcompanies has already applied NLP techniques to create virtual assistant likeSiri (Apple), Cortana (Microsoft), Alexa (Amazon) etc. This paper representsideas for further developing virtual assistants so they can be used in dailylife. Some deep learning can be introduced in natural language processing whichmay help us build a more advanced virtual assistant.

INTRODUCTIONMachine learning and deep learning  fueled by big data are helping mankind tosolve intelligence and companies like DeepMind and OpenAI are pushing theboundaries of AI to make it more practical and usable in our day to day life.One Such application is creating virtual assistants that requires concepts fromdisciplines like NLP and Deep Learning. One such conceptis recursive neural network as mentioned in 1,it takes and understands ininformation by the previous stored information in its hidden layer which is offixed size.LITERATURE SURVEYLinguistic and computational knowledge is needed  for NLP.

while extracting the informationthere are many problems like paraphrasing,understanding idioms and metaphors.Byintroduncing the concepts of deep learning in NLP the machine will be able tounderstand the language better like use of adjectives and sarcastic sentences.The2 major classifications of approaches for NLP 1 are:1. Statistical learning approach     2. Rule-based approach Since textual tree construction can be time consuming forlong sentences, it is inefficient. Recurrent Neural Network can extractcontextual information by utilizing stored previous text in the form of fixedsized hidden layer.

This deeper learning in language processing will bebeneficial for further developments of virtual assistants.The idea of  use virtual assistants for self driven carshave also been proposed2,embedding autonomy,deep learning and adaptabilitycan help improve and drive better than humans.The human brain of drivers isemotionally aware and developed and it takes decisions according to the gutfeeling and past experiences but this is not so in machines and virtualassistants.The combination of artificial neural networks and deep learning willsimplify the task of replicating the functioning of human brain in machines andmake the artificial brains emotionally aware.The artificial brain of virtualassistants should be able to take independent decisions,obey trafficlaws,optimize fuel consumption,reduce pollution and ensure better safetyfeatures for the passengers.In the workshop “talking with conversational agentsin collaborative action” as mentioned in 2 in mobile phones the evolutionfrom touch to speech interface idea was proposed with smart T.V.s,smart watchesetc.

which are now have been successfully implemented.In virtual assistantthere is typical use of speech based queries.The idea of”virtual Buttler” wasalso proposed in a workshop which will act like a real person and practices longterm  use of a companion.

Text to speechand speech to text are very popular applications of NLP used in creatingchatbots.The idea of expressive virtual text to speech system was presentedusing deep neural network 3.In this a text sentence will be given with aexpressive tag which will be used to produce a photorealistic talking head.Thespeaker adaption technique uses small training data from a novel speaker tomodify a system that has already been trained.The system can be modified inthree ways3:1.Speaker specific code is appended to input 2.

Speakerspecific reweighing of hidden contributions is learnt  3.Speaker dependent mapping is learnt fromeach speaker.In the visual model of virtual assistant image of face isconstructed by separating modes into semantically meaningful actions andregions.One Mode is used for model blinking,two modes are used for 3D headrotation,eight modes are used to model lower half of the face and six modes areused to model upper half of the face. There are expression space finalexpression dependent layers for mapping to specific expressions already codedcalled “Expression Adaptation”3.Some new expressions can also be added to themodel with small amount of adaptation data and new output layer can be addedfor  the same.

The linguistic feature issent to the penultimate layer of previously trained Deep Neural Network forproducing final expressions in the virtual head model.Regularised form oflinear least square algorithm should be used to avoid overfitting on smallamount of adaptation data.Different layers are formed for different expressionslike anger,fear.happy,sad and tender.The vehicles use sensors for vehicle tovehicle communication.

When a lead vehicle suddenly deaccelerates a MATLABsimulink model varies vehicle dynamics,following vehicle’s distance and initialspeeds.Artificial Intelligence can improve cases of escaping injury,vehicledamage and traffic jams.The driving assistant uses telematics,globalpositioning and sensors ,it can avoid potential dangers.The warning methodologyis constructed through  “spatio temporalsafety zones”4,the three zones are:1.Accident Mitigation (AM)  2.

Accident Avoidance(AA)  3.Accident Free(AF)After examining these three zones the system increases theradius of the vehicle concentrically from the vehicle’s center. Each vehiclemonitors other vehicles constantly in case the vehicle assistant needs toreact.Feed forward neural network approach is used  if any vehicle suddenly applies brakes infront of the vehicle following it.

Another application of virtual assistant hasbeen proposed in the field of medicine “The Mindbot” 5 for mental healthproblems.Psychotherapies have been developed to identify facts  that cause mental illness.This virtual agentwill behave as a real human therapist and will ask mental health relatedquestions from the patients and the answers will be evaluated according to thestandard psychological scales and results will show the mental fitness level.Theconcept of “Listen while speaking” 6 for virtual assistant in all the devicescan soon be introduced by the use of peech chain in deep learning.The speechchain is a process that helps replicating the exact human speech using speechchain algorithm as referred in 6.The technology of automatic speechrecognition (ASR)and text to speech recognition(TTS) enables machine to processand respond to human speech.FUTURE SCOPEThe use of virtual assistants is gaining a lot of popularityin many fields and the interest in deep learning.In future the virtualassistants will help humans in their day to day life.We will be able to embedvirtual  assistants in many public placeslike railway stations,hospitals, banks.Installing virtual assistants in ATMS willhelp  visually challenged people toaccess the facility easily without being dependent on theIrs ans we can furtherwiden their usability by introducing virtual assistants in different languagesincluding the local languages of different states of India.