However if I have to change the address in my insurance, or any other service provider’s records, I can’t simply send an email, and expect a useful response immediately. What if I get sick, and need to lookup nearby hospitals that are covered by my insurance, an email to my insurance company would never elicit a quick response. Most often than not, I will receive a dummy message that will read, “Your enquiry has been recorded. We will respond within 48 hours.”
Why should it take this long to get me this simple information? Technology available today should be able to get this information within seconds. What is needed is an AI based auto responder, and not the archaic dummy from the past.
We at Flowmagic have always maintained that human ingenuity and creative skills should be used for innovative things. There is no big deal in answering the same questions that a person would have answered thousand times over and over again. From that perspective, any repetitive jobs that can be done with Artificial Intelligence should be relegated to it. AI Autoresponders work will free up your agents time for more complex problems and their resolutions. Agents will also be relieved of having to answer repeated questions from various customers, because the same can be achieved with AI Autoresponders.
FlowMagic AI Email Responder
FlowMagic AI Email Responders studies, and analyses the email subject, body, and the context of the sender of the message. Specific information provided in the mail is understood in real time, and an appropriate response created on the fly. FlowMagic also customises its messages based on the context of the historical messages sent by the customer over a period of time.
How does it work?
At the beginning of AI Auto Responder, data is preprocessed in the following way:
• Language Detection: The language of the email is detected, and non english messages are routed for manual response.
• Tokenization: Subject and mail body are broken into words and punctuation marks
• Sentence Segmentation: Sentence boundaries are identified in the message body.
• Normalization: Infrequent words are replaced with special tokens.
• Quotation Removal: Quoted original messages are removed.
• Salutation Removal: Salutations like Hi, Dear, etc are removed.
This AI Email Responder has been trained on millions of emails messages, and customer support answers to these emails. The responder uses Sequence based Deep Neural LSTM Networks with multiple knowledge graphs to ensure robustness, and accuracy, and most importantly usefulness of the responses.
Training & Evaluation
Dense Vectors vs One Hot Representations
A modern day autoresponder is unlike the linear models of SVM, or Logistical Regression. Traditional methods generally used one hot feature vector which results in sparse feature vectors. Instead Sequence based Recurrent Neural Network do not use sparse one hot vector approach, but use dense word embedding vectors. That is each core feature is embedded into a d dimensional space, and represented as a vector in that space. The embeddings (the vector representation of each core feature) is then trained like any other parameter in a neural network. One benefit of using dense and low dimensional vectors is computational: the majority of neural network toolkits do not play well with very high dimensional, sparse vectors. However, this is just a technical obstacle, which can be resolved with some engineering effort. The main benefit of dense representation is in generalisation power: if we believe that some features may provide similar clues, it is worthwhile to provide a representation that is able to capture these similarities.
Many bots follow a predefined path and are not very capable of answering free flowing messages from the user. AI Auto Responder’s NLP capabilities can handle these messages with equal ease, and can suitably respond