Consumers really don’t like your chatbot. It’s not exactly a relationship built to last — a few clicks here, a few sentences there — but Forrester Analytics data shows us very clearly that, to consumers, your chatbot isn’t exactly “swipe right” material. That’s unfortunate, because using a chatbot for customer service can be incredibly effective when done […]
“HubSpot's GrowthBot is an all-in-one chatbot which helps marketers and sales people be more productive by providing access to relevant data and services using a conversational interface. With GrowthBot, marketers can get help creating content, researching competitors, and monitoring their analytics. Through Amazon Lex, we're adding sophisticated natural language processing capabilities that helps GrowthBot provide a more intuitive UI for our users. Amazon Lex lets us take advantage of advanced AI and machine learning without having to code the algorithms ourselves.”
There was a time when even some of the most prominent minds believed that a machine could not be as intelligent as humans but in 1991, the start of the Loebner Prize competitions began to prove otherwise. The competition awards the best performing chatbot that convinces the judges that it is some form of intelligence. But despite the tremendous development of chatbots and their ability to execute intelligent behavior not displayed by humans, chatbots still do not have the accuracy to understand the context of questions in every situation each time.
Your first question is how much of it does she want? 1 litre? 500ml? 200? She tells you she wants a 1 litre Tropicana 100% Orange Juice. Now you know that regular Tropicana is easily available, but 100% is hard to come by, so you call up a few stores beforehand to see where it’s available. You find one store that’s pretty close by, so you go back to your mother and tell her you found what she wanted. It’s $2, maybe $3, and after asking her for the money, you go on your way.
I will not go into the details of extracting each feature value here. It can be referred from the documentation of rasa-core link that I provided above. So, assuming we extracted all the required feature values from the sample conversations in the required format, we can then train an AI model like LSTM followed by softmax to predict the next_action. Referring to the above figure, this is what the ‘dialogue management’ component does. Why LSTM is more appropriate? — As mentioned above, we want our model to be context aware and look back into the conversational history to predict the next_action. This is akin to a time-series model (pls see my other LSTM-Time series article) and hence can be best captured in the memory state of the LSTM model. The amount of conversational history we want to look back can be a configurable hyper-parameter to the model.
Getting the remaining values (information that user would have provided to bot’s previous questions, bot’s previous action, results of the API call etc.,) is little bit tricky and here is where the dialogue manager component takes over. These feature values will need to be extracted from the training data that the user will define in the form of sample conversations between the user and the bot. These sample conversations should be prepared in such a fashion that they capture most of the possible conversational flows while pretending to be both an user and a bot.
Note — If the plan is to build the sample conversations from the scratch, then one recommended way is to use an approach called interactive learning. We will not go into the details of the interactive learning here, but to put it in simple terms and as the name suggests, it is a user interface application that will prompt the user to input the user request and then the dialogue manager model will come up with its top choices for predicting the best next_action, prompting the user again to confirm on its priority of learned choices. The model uses this feedback to refine its predictions for next time (This is like a reinforcement learning technique wherein the model is rewarded for its correct predictions).
In a particularly alarming example of unexpected consequences, the bots soon began to devise their own language – in a sense. After being online for a short time, researchers discovered that their bots had begun to deviate significantly from pre-programmed conversational pathways and were responding to users (and each other) in an increasingly strange way, ultimately creating their own language without any human input.
If a text-sending algorithm can pass itself off as a human instead of a chatbot, its message would be more credible. Therefore, human-seeming chatbots with well-crafted online identities could start scattering fake news that seem plausible, for instance making false claims during a presidential election. With enough chatbots, it might be even possible to achieve artificial social proof.
In so many ways I think chatbots are only just getting started – their potential is much underestimated at present. A big challenge is for chatbots mature so that they do more than is possible as a result of content entry wizards. If your content is created with a few easy clicks, it is unlikely to be much inspiration to anyone – and to date, despite much work in the field, the ability to emulated the creative open ended nature of real intellingence has seen only very partial success.
The bot (which also offers users the opportunity to chat with your friendly neighborhood Spiderman) isn’t a true conversational agent, in the sense that the bot’s responses are currently a little limited; this isn’t a truly “freestyle” chatbot. For example, in the conversation above, the bot didn’t recognize the reply as a valid response – kind of a bummer if you’re hoping for an immersive experience.
Lack contextual awareness. Not everyone has all of the data that Google has – but chatbots today lack the awareness that we expect them to have. We assume that chatbot technology will know our IP address, browsing history, previous purchases, but that is just not the case today. I would argue that many chatbots even lack basic connection to other data silos to improve their ability to answer questions.
The term "ChatterBot" was originally coined by Michael Mauldin (creator of the first Verbot, Julia) in 1994 to describe these conversational programs. Today, most chatbots are accessed via virtual assistants such as Google Assistant and Amazon Alexa, via messaging apps such as Facebook Messenger or WeChat, or via individual organizations' apps and websites. Chatbots can be classified into usage categories such as conversational commerce (e-commerce via chat), analytics, communication, customer support, design, developer tools, education, entertainment, finance, food, games, health, HR, marketing, news, personal, productivity, shopping, social, sports, travel and utilities.