These days, checking the headlines over morning coffee is as much about figuring out if we should be hunkering down in the basement preparing for imminent nuclear annihilation as it is about keeping up with the day’s headlines. Unfortunately, even the most diligent newshounds may find it difficult to distinguish the signal from the noise, which is why NBC launched its NBC Politics Bot on Facebook Messenger shortly before the U.S. presidential election in 2016.
There is a general worry that the bot can’t understand the intent of the customer. The bots are first trained with the actual data. Most companies that already have a chatbot must be having logs of conversations. Developers use that logs to analyze what customers are trying to ask and what does that mean. With a combination of Machine Learning models and tools built, developers match questions that customer asks and answers with the best suitable answer. For example: If a customer is asking “Where is my payment receipt?” and “I have not received a payment receipt”, mean the same thing. Developers strength is in training the models so that the chatbot is able to connect both of those questions to correct intent and as an output produces the correct answer. If there is no extensive data available, different APIs data can be used to train the chatbot.
It won’t be an easy march though once we get to the nitty-gritty details. For example, I heard through the grapevine that when Starbucks looked at the voice data they collected from customer orders, they found that there are a few millions unique ways to order. (For those in the field, I’m talking about unique user utterances.) This is to be expected given the wild combinations of latte vs mocha, dairy vs soy, grande vs trenta, extra-hot vs iced, room vs no-room, for here vs to-go, snack variety, spoken accent diversity, etc. The AI practitioner will soon curse all these dimensions before taking a deep learning breath and getting to work. I feel though that given practically unlimited data, deep learning is now good enough to overcome this problem, and it is only a matter of couple of years until we see these TODA solutions deployed. One technique to watch is Generative Adversarial Nets (GAN). Roughly speaking, GAN engages itself in an iterative game of counterfeiting real stuffs, getting caught by the police neural network, improving counterfeiting skill, and rinse-and-repeating until it can pass as your Starbucks’ order-taking person, given enough data and iterations.
For each kind of question, a unique pattern must be available in the database to provide a suitable response. With lots of combination on patterns, it creates a hierarchical structure. We use algorithms to reduce the classifiers and generate the more manageable structure. Computer scientists call it a “Reductionist” approach- in order to give a simplified solution, it reduces the problem.
These are hardly ideas of Hollywood’s science fiction. Even when the Starbucks bot can sound like Scarlett Johansson’s Samantha, the public will be unimpressed — we would prefer a real human interaction. Yet the public won’t have a choice; efficient task-oriented dialog agents will be the automatic vending machines and airport check-in kiosks of the near future.
The goal of intent-based bots is to solve user queries on a one to one basis. With each question answered it can adapt to the user behavior. The more data the bots receive, the more intelligent they become. Great examples of intent-based bots are Siri, Google Assistant, and Amazon Alexa. The bot has the ability to extract contextual information such as location, and state information like chat history, to suggest appropriate solutions in a specific situation.
However, as irresistible as this story was to news outlets, Facebook’s engineers didn’t pull the plug on the experiment out of fear the bots were somehow secretly colluding to usurp their meatbag overlords and usher in a new age of machine dominance. They ended the experiment due to the fact that, once the bots had deviated far enough from acceptable English language parameters, the data gleaned by the conversational aspects of the test was of limited value.
One pertinent field of AI research is natural language processing. Usually, weak AI fields employ specialized software or programming languages created specifically for the narrow function required. For example, A.L.I.C.E. uses a markup language called AIML, which is specific to its function as a conversational agent, and has since been adopted by various other developers of, so called, Alicebots. Nevertheless, A.L.I.C.E. is still purely based on pattern matching techniques without any reasoning capabilities, the same technique ELIZA was using back in 1966. This is not strong AI, which would require sapience and logical reasoning abilities.