Designing Personality: How Bots Express Themselves Through Language
Developers and companies are building bots to sound more like humans, Using these personality design components can improve user experience with your chatbots.
August 7, 2018
What makes a personality? Consistency, interests, and a distinct voice are a few key aspects. The ability to understand responses and to move the conversation forward is, for chatbots, or bots, an essential part of communicating like a human. With recent steps forward in deep learning, we’re entering a phase of more intelligent bots, but we’re not quite at a point where they come across as human. (For instance, no bot has yet passed the Turing test.) As companies increasingly use bots for customer service or as personal assistants, and as bots provide people with AI friends (like Mitsuku), making sure that bots can effectively express themselves in ways that humans recognize will become even more important. Let’s take a closer look at how the industry is creating expressive bots and what steps companies can make in this direction.
Give bots human communication skills
There’s a reason that communication is listed as a required skill in almost every job description. Verbal communication is one of the primary ways that humans connect and, if we want bots that seem human, we need to equip them accordingly. Incorporating linguistic knowledge in bot training will help get us there. Linguistic skills, such as semantics, syntax, and pragmatics, which are aspects of NLP, will make bots seem more human. The hardest of these is pragmatics, which is about understanding not only what is said but also what is meant. The industry hasn’t quite figured out pragmatic analysis yet, but teaching bots to ask clarifying questions or to rephrase user inputs can help bridge the gap.
One of the most human communication skills is the ability to improvise in conversation. Improvisation creates a sense of thought behind the speech, fuels small talk, and moves conversations forward. ELIZA simply formulates a question based on what a human says to her, but more modern bots must do more than parrot back a user’s inputs; they need to be able to offer fresh thoughts.
Facebook was exploring this in 2017 when it was trying to teach bots how to negotiate—by having them talk to each other. Bots Alice and Bob were instructed to negotiate with each other, and they quickly started talking in what’s been called a coded language, which did not make sense to their human developers. Although this experiment may not have been a success, it was evidence that bots can “develop their own languages, including languages with a coherent structure, and defined vocabulary and syntax.” That’s a big step toward being able to improvise using human speech. Facebook is currently tackling this with a new dataset, called Persona-Chat, trying to teach bots how to make small talk.
Teach bots how to feel
Other skills that are important for human-seeming bots are empathy and the ability to vary responses. Empathy in a bot isn’t a new concept—in fact, ELIZA modeled this. Newer bots Woebot and Wysa are both designed to provide empathetic, therapeutic conversation. But even bots that aren’t designed for therapy need to show empathy, because human speech is infused with feelings.
Response variation is also key and can go hand-in-hand with using empathy to foster connection. Humans say things in different ways. This is, in fact, a necessary skill in furthering understanding between two interlocutors. Rephrasing what someone has said, using your own words to confirm you understand their meaning, is an active listening skill that helps the speaker feel understood. A bot that repeats itself conveys that it does not understand what a user is trying to say sounds mechanical rather than human. IBM’s Watson, for instance, can offer different responses that have similar intent.
Control the flow of conversation
Typical conversational elements, such as pauses, interjections (“oh,” “hmm,” and the like), and response times can go a long way toward giving bots a human feel. Alfred, a newer bot from IBM, designed to help project managers spend less time on administrative tasks, uses interjections to create a connection.
Bots may be able to respond lightning-fast to user inputs, but building in pauses can help them feel more human. Rose, a two-time Loebner Prize winner, provides a good example of personality. Her persona is that of a computer nerd with a quirky attitude who is from an unorthodox family. She used to be a computer security analyst and is now moving into AI. She enjoys battling robots and playing ARGs, and she cracks jokes about politics.
But as advanced as Rose is, her split-second responses are disorienting and actually make it less clear which comment she’s responding to. Of course, a pause that’s too lengthy between question and response won't go over well, either. Evaluating typical response times in human-to-human, real-time digital interactions can help bot developers improve this detail.
Make bots funny
Humor is a cornerstone of human conversation. Companion bots, like Rose and Mitsuku, are good at this, but developers can build humor into any type of bot to give them a more human feel. People like others who make them laugh and even view them as more intelligent. When it comes to bots, humor can help foster an initial connection, alleviate tension, and keep users engaged while a bot works on completing a task.
Debt Like WTF, a bot designed to help people save on student loans, uses humor at the beginning of an interaction to build connection. Although some bots can succeed with cracking jokes up front, for the most part, “humor is something that should organically unfold instead of seeming overly forced,” according to Emily Withrow, an editor at Quartz Bot Studio.
Remember personal information
When talking to a person, you naturally expect that they’ll remember what you’ve told them, whether earlier in the conversation or on a different day. When a listener remembers what you’ve said, it signals that they’re paying attention. A demonstrated awareness of having interacted with a user in the past, or a recollection of something said earlier in the conversation that provides context, will help bots better connect with humans.
Even a simple bot designed to help someone order a pizza will benefit from remembering what that person has ordered in the past. Such a bot can ask if the user wants the same thing as last time or can make suggestions based on known preferences. To be effective, task-oriented bots need to help users save time. And, given the availability of user data, bots should be able to demonstrate the same kind of memory that an acquaintance can offer.
Developers are making progress in this area. In 2016, two students launched Wonder, a bot designed to remember user information and to provide it when asked. Facebook has been experimenting with memory networks for years and, in addition to making bots more useful, the improved memory will make them seem more human.
The problem of datasets
Lack of robust datasets continues to be a challenge in training AI-based bots. Aside from employing writers to create natural conversational scripts, the best way to teach a bot to speak like a human is to use real-life human interactions and language. Training data for customer service bots can come from company resources as well as from various publicly available archives or movie scripts. However, for increasingly advanced personal assistant or companion bots, these sources are often too stiff or too limited to provide good training data.
Big companies are moving forward in this territory by turning to alternative sources. Facebook’s Persona-Chat has 160,000 lines of dialogue sourced from Mechanical Turk. At Microsoft, researchers are using dialogue from Twitter and Reddit to provide data truer to typical human interactions. Of course, this is a risky business, as Microsoft’s history with socially offensive bots has shown. To prevent inappropriate or prejudicial responses, developers will need to continue to adjust algorithms and should expect to make mistakes. Further advances in sourcing and filtering data will help bots sound more human and exhibit more personality.
Bot designers are starting to focus more on bot conversation and personality. Along with constructing personality, developers can give bots a more human feel. NLP’s lack of pragmatic analysis skill, along with the shortage of datasets (in addition to the poor quality of those that exist), continues to limit how bots can seem human. However, as these disciplines progress and converge with changes in UX design, bots will be able to sound increasingly human.