In our previous article on chatbots, it was stated that one does not simply create a chatbot. Artificial intelligence still poses too many challenges to go about building bots without sufficient research. One of the main problems is the ability of bots to recognize human requests. We’ve decided to tell you about some of the popular AI chatbots that handle this task well.
Why research matters
In theory, the best chat bots were made to replace human services. Why rely on emotional creatures that get tired when you can get help from a talking database? In practice, bots are truly good at retrieving information. What they not so good at is understanding what you want.
The ability to get the context is called natural language understanding, or NLU. NLU is considered an AI-hard problem, meaning that a part of the process should be still done by the humans. A bot called Rose (developed by Bruce Wilcox), the 2015 winner of Loebner Prize, is so highly appreciated because it recognizes idioms and initiates conversations. This ability is a result of a prolonged testing and considerable human input.
Below are some of the more successful examples from expert ratings. In this article, we’ll focus on what it takes to make every artificial intelligence chatbot on the list understand human language.
A ‘mother’ of self-learning chatbots, Apple’s Siri was launched in 2011. It was the first scalable assistant with recognition of speech and ability to learn by observing users.
For a 2-year-old startup, it was quite a challenge to bring together several technologies:
- local search engine;
- AI technologies;
- special data processing and storage systems.
But this was only the end of a long process. Developing the entire concept took Adam Cheyer, co-creator of Siri, almost 20 years.
Cheyer once told The Startup that the hardest part of the process was dealing with the ambiguity of human language. Any word can be a business name; some names, even city names, are identical across different territories.
Most of us know Google’s virtual assistant that handles search, home chores, schedule and written interactions. Its new version lets users query using both voice and text. It also boasts rich functionality for mobile shopping. However, less is known about how the Assistant has been built.
Just like Siri, Google Assistant has its own unique personality. The company has a separate team working on it, headed by ex-Pixar author Emma Coats. According to Wired, this department works in iterations. It comes up with sets of questions and creates answers, often humorous. The answers are then handed down to the developers team.
The team admits that Google Assistant is learning to function without human help. The algorithm collects human requests and reacts accordingly. Machine learning is one of the latest AI chatbot trends, and Google seems to follow suit. Fernando Pereira, Google’s head of NLU projects, claims that the bot will soon be implicitly learning rather than taught. This set of ideas and practices is often referred to as ‘The Transition.’
Alexa is a smart home virtual agent. Unlike other voice recognition solutions, this one is only available through Amazon devices such as Echo. Amazon lets third-party developers add ‘skills’ (services that work with the platform).
In November 2016, the company announced that they would make Alexa’s voice recognition technology open. Amazon employs several teams to design and develop this solution. The company has even dedicated a separate conference entirely to Alexa.
The graph below is a takeaway from the benchmark of Google API.ai, LUIS, Alexa, and other NLU systems by Caroline Wisniewski, Clément Delpuech, David Leroy, François Pivan, and Joseph Dureau (visualized with Piktochart).
This artificial intelligence chatbot was launched in December 2016 to help the clients of Royal Bank of Scotland (RBS). Nexus team is reported to have been creating Luvo’s personality for nearly half a year. A big amount of this work related to language processing: the appropriate formulas and empathy. All of this has been done before they started the actual coding. Paying attention to the bot’s personality is one of the main company’s tips for anyone willing to create AI-based tech.
Xiaoice (earlier Xiaobing) is a Mandarin-language extension of Microsoft’s Cortana developed exclusively for relationship-building. Like all of the best artificial intelligence chatbots, the solution combines machine learning with big data and linguistic analysis.
The bot learns to behave as a cute teenage girl. It is using smileys and emojis perfectly. Xiaoice’s success could be explained by the fact that Microsoft researchers work on both her IQ and EQ. The team working on the bot includes psychologists who have been developing a set of scenarios and compassionate questions. Considerable memory is another cornerstone of the success.
Lark is a personal fitness tracker and healthcare coach available for iOS and Android. The bot handles several tasks:
- asks users about their daily habits;
- gets data from fitness trackers;
- gives them custom answers from the database of expert advices.
Google called Lark the 2016 best AI chatbot app. This result did not come out of the blue: it was preceded by a 6-year research with health and behavior change experts from Harvard and Stanford, as well as artificial intelligence experts.
Another popular healthcare chatbot is currently being tested by NHS.
What? Poncho is on the list of best AI bots for Messenger? You’d say that we must be joking. But we are not: there is a way in which this young and inexperienced weather assistant outsmarts many other bots. It has a mechanism to protect itself while actively learning from users.
Poncho CEO Sam Mandel told the media that the bot has a kind of internal score for each user. When someone behaves inappropriately, Poncho asks to apologize first and then remains silent for 24 hours. This way, the technology protects itself from learning inappropriate behavior, the grim fate of Microsoft’s self-learning bot called Tay.
Launched in 2015, Mezi quickly became a popular shopping assistant for busy people. It is taught to filter out important information from spoken requests. With this filtered request, Mezi proceeds to choose the best goods at the most favorable price.
Mezi’s AI is called Smart Assist. At the moment of launch, it was responsible for 20% of the app reactions while 80% of it relied on human intelligence. The startup hopes to flip this ratio gradually. As for 2017, according to Mezi co-founder Shehal Shinde, the staff helped the app’s AI when there was a big traffic breakthrough.
While Lark operates as a health coach, this chatbot gives automation tools to recruiters:
- filters job applications;
- provides consultations on company policies and culture;
- organizes the resumes that it picked into a comprehensive list.
The problem with bots like Mya is that they can unwittingly learn human stereotypes. Say, it can favor male candidates over female due to language bias. Once more, this proves that AI is impossible without a touch of humanity.
The graph below (brought to you by Softermii, powered by Piktochart) shows how much R&D work is needed to launch an innovative chatbot. Artificial intelligence at its early stages still needs human efforts to grow.