Your Complete Guide To Chatbots
Chatbots are generally computer programs that can interact with a user. They deliver natural, human like responses in a conversational style, usually as a substitute to a real person. This provides quicker responses to customer needs at lower cost. They usually use text to converse, but increasingly verbal communication is also possible. A chatbot can talk to hundreds of people at a time, whereas a human can typically only have a meaningful chat session with two or possibly three at once. These ‘bots’ are often found where you might otherwise see live chat offered on a website, typically a bubble or tab in the bottom right corner of a web site, oron other chat / instant messaging platforms such as Facebook Messenger or Skype.
Everyone seems to be using chatbots these days. Artificial Intelligence is undergoing a major resurgence, and for good reason - although it is far from new. Computational power and the advent of cloud computing has brought it to the fore, but systems which can perform reasoning tasks, speech generation and language processing go back several decades.
The first experimental chatbots were designed to explore both how humans produce and understand language and how effectively a computer could process natural language, effectively putting different methods of language analysis and production to the test. Computer scientist Alan Turing, in an attempt to answer the question ‘can machines think?’, devised the Turing test; developed in 1950, the test was to determine whether one could distinguish between a human and a computer just through textual conversation.
While some chatbots can be quite sophisticated they generally do not pass the Turing test as they cannot maintain an illusion of understanding for very long.
Some chatbots created in the 1960s and 70s were quite convincing on first impression. One such program running on a TRS-80 - a machine with minimal computational power or storage by today’s standards - was quite impressive, able to fool users into believing they were speaking with something which understood them. The intelligence was largely illusory, but convincing - in part because the concept of Artificial Intelligence was not widely known at the time and therefore not understood by the average user. Since then, chatbots have moved on significantly both in computational power and technique; the most notable development being the use of machine or ‘deep’ learning. Deep learning, using a kind of artificial neural network which mimics some aspects of the brain’s structure, allows a system to learn from examples, rather than requiring every possible input and output to be explicitly coded. This is an improvement on the older chatbots, but it is definitely not a panacea for creating synthetic minds, and these newer, shinier chatbots come with their own negative points, too.
Chatbots can be driven by voice, text, mouse or gesture.Text chat bots have the advantage of being accessible through messaging systems such as Facebook messenger, Skype, WhatsApp or other platforms, allowing them to be available almost anywhere.
The synthetic speech makes interactions appear more natural. Even your average SatNav can have a natural, clear voice and we think nothing of taking instructions from it! Speech recognition, as opposed to speech generation, now allows us to engage in two-way spoken dialogue. Speech recognition is impressive, from a technology standpoint, but it still has a little way to go to be sufficiently reliable. The system can misinterpret what is being said, resulting in repeat attempts and frustration, particularly in a noisy environment or exposure to a different accent – the classic humorous example being someone with a thick Scottish accent stuck in a voice operated lift/elevator with no means of escape!
Nonetheless, we are now able to communicate with devices purely by voice and, thanks to home-based devices like Amazon Alexa or GoogleHome, the technology is constantly improving. Increased usage translates into a huge spike in learning opportunities for the chatbot, due to increased exposure to real human dialogue patterns and increased practice at delivering suitable responses.
More recent uses of the term “chatbot” have expanded to include variations which do not require natural language processing at all. For example, some chatbots are designed to recognise a selected pre-provided answer from a set list. While this can be effective for simple tasks, it does not give the user the freedom to control the conversation or provide information voluntarily. This can be a limitation but, at the same time, as the chatbot has complete control over the flow of the dialogue, the results are more predictable and more easily controlled. Whether such systems can truly be classified as chatbots is debateable, but they can perform useful functions nonetheless.
These multiple methods of input provide great flexibility and give the appearance of some underlying intelligence which can be impressive. But this apparent intelligence can be deceiving. Are these chatbots truly smart enough to replace real people?
Many chatbot systems on the market seem to focus on the communication channel, the speech, or the web interaction - in otherwords, what it looks and sounds like, as opposed to what it can actually do and say. A flashy profile combined with vague references to “machine learning” may convince you that you are some how buying into an intelligence system for just a few tens of dollars a month. You would be right to be sceptical.
With some rehearsed demonstrations, using quality speech generation and recognition, it’s easy to be fooled into thinking you are speaking with an intelligent system, or even a synthetic mind. But without the right encoded knowledge, the ability to respond to conversational cues, and without understanding the context and purpose of the conversation, the system is an impressive illusion.
Conversely some chatbot marketing material you may find aims to impress you by focusing on particular examples where simpler solutions would be fooled. There are some things which humans find easy which chatbots typically find difficult to understand. Regional expressions, colloquialisms or simply ambiguous text can easily confuse a chatbot. Where a real person would understand from the context, or seek toclarify, a chatbot makes assumptions.
Solving this kind of problem requires real world knowledge and is not easily accomplished. With a large enough data set, this could potentially be handled using Machine Learning techniques. Machine Learning requires large amounts of data, usually tagged or categorised in someway by humans, and the data must be free from unintended biases. There sulting system can then respond in an impressive manner, but it can also be fooled or generate non sensical responses. The nature of machine learning and neural nets make it impossible to determine how a particular conclusion was drawn. We can only guess how a chatbot learned a specific response. So, if a chatbot is producing unsuitable responses, the only recourse is to retrain it with more or different data examples. Meanwhile, unsuitable responses which go undetected could have significant consequences to a business or a person, if not identified as mistakes in a timely manner. Mistakes are normally handled by adding appropriate confirmation statements and requests to ensure that the system fully understood the user input.
Knowing the topic of conversation beforehand offers a significant advantage. Direct methods of addressing mistakes can be employed, using our own understanding of grammar and vocabulary. This, which is sometimes referred to as symbolic reasoning, can still be challenging and may require many cycles of testing and improvement as feedback is acquired over time.
Then again, depending on the nature of the chatbot, complex language processing may not be necessary at all. It depends on what the bot is for and the scope of conversation it is intended to handle. If the chatbot is to perform psychotherapy for patients, then it will need to be very sophisticated. If it is intended to assist with product selection, completing online orders or booking flights then it can be comparatively simple, aslong as it requests appropriate confirmations before taking action. If the chatbot is intended to resolve problems with faulty cookers, help users operate complex systems, or find your perfect holiday, then it’s probably going to be somewhere in between.
When considering any investment in a chatbot there are several things to consider.
For a simple chatbot, there are numerous online offerings which can perform simple tasks such as gathering a prospect’s details or booking an appointment. These can operate through many of the channels we have discussed and can be set up relatively quickly, at a moderate ongoing cost. The chatbot’s intended use will be key to anticipating how easy it will be to create and develop and how readily it can be customised. While having a chatbot fit in with your brand’s font and colours may be important, if your chatbot doesn’t have anything useful to say it will not get very far. Any system can look good, but it takes skill to build a chatbot that can generate convincing responses to a real human request. Systems with simple point-and-click interfaces will be easier toset up but will not have the flexibility of free-form text input and cannot be accessed through messaging or voice channels.
Looking at the conversational editing capabilities will be important – much more important than how it looks or sounds. A great-looking chatbot with no conversational skills will only disappoint. For standard cloud-based offerings the editing capabilities consist of a number of sections for recognising the input or ‘utterances’, providing the necessary responses that the chatbot can respond with, and the means by which information can be extracted from the input voice or text specific to the customers’ requests or responses. Many systems offer free trials, and it is important to see whether the conversational creation tools are going tobe up to the task.
Some chatbots, even those created by well-recognised global organisations, can fall short when the system’s responses must go beyond the trivial. If the chatbot’s required responses are well-targeted and designed to perform only a few specific functions, then a pre-made bot may well be sufficient.
For a chatbot to develop and improve there needs to be a means by which its performance can be measured, and gaps in its knowledge identified and improved. User-entered content which has not been recognised by the chatbot needs to be identified and either associated with the systems existing conversational drivers or new conversational flows which have not yet been developed.
A chatbot is never able to handle all possible conversational scenarios, much the same as their human counterparts. It is therefore important for the system to recognise when it is out of its depth and provide an alternative means to assist with the customer’s enquiry, typically by logging the information acquired and arranging for a human to intervene or take over.
You should consider custom chatbot development to ensure you have the flexibility you need. Standard cloud offerings do not have the fullflexibility of a custom-built solution. Cloud options will provide some flexibility but will usually be restricted in how they are presented or what channels they can use. They are also somewhat restricted on how the knowledge content is created, relying on the manual input of thousands of examples - or endless additional ‘flows’ - to improve its ability to respond effectively.
Data from your business can be used to help drive the dialogue - this may be product or service information, historical events, previous interactions with online chat agents, or other information about your customers which can be used to ensure the conversation is personalised for each individual and their needs. With a custom solution this can be built into the conversation engine and administrative interfaces in the way that works best for you.
Creating a custom solution means considering your specific requirements and aspirations and finding the best way to fulfil these. We can help bring the necessary components together into a specialised solution which is just right for you. This may involve using existing cloud or open source chatbot or AI services and integrating them together into a holistic solution. Alternatively, you may already have created a chatbot in some form and found that you have reached the limits of the solution’s capabilities. We can help you find a way forward which doesn’t mean starting again from scratch.
For custom-made chatbots that are tailored to you and your business, call Brandon Cross on 020 8144 2000.