Why smart conversation designers study linguistics
Conversation design is about creating natural and human-like conversations. As a conversation designer, your design choices are based on informed decisions — not a matter of taste or an expression of personal style. ‘I like it,’ is not a particularly great argument, whereas ‘this is proven to work’ is a far better explanation.
That’s why smart conversation designers study linguistics.
Grice’s maxims
To create a good conversation, we need to understand what the rules of good conversation are. Luckily, the scientific study of language provides us with very concrete insights into the inner workings.
By now, Paul Grice’s Cooperative Principle is well known. It describes how conversational partners help each other out during a conversation, often unconsciously. As people, we try to make our exchanges as effective and productive as possible for both parties. We infer meaning, skip steps, establish common ground and build rapport with our conversational partner.
The four maxims are:
- The maxim of quantity, where one tries to be as informative as can be, not give less information than is needed, and no more.
- The maxim of quality, where one tries to be truthful and not give false information.
- The maxim of relation, where one tries to be relevant to the context of the conversation.
- The maxim of manner, when one tries to be as clear and brief as possible, avoiding obscurity and ambiguity.
Grice’s maxims are to be understood as describing the assumptions listeners make about the way we talk, rather than prescriptions for how one ought to talk.
However, the beauty of Grice’s maxims is that they were written down in the form of guidelines. It makes them easily applicable to conversation design. Using these guidelines we can see if we’ve taken into account the common assumptions that listeners make when designing conversations.
As a result, users will be able to communicate with your bot as if they were talking to a friend, colleague or someone on the phone.
Conversational implicatures
According to Paul Grice, one of the biggest challenges in understanding speech lies in the difference between what is said and what is meant. People infer meaning. If intended by the speaker, these additional meanings are called conversational implicatures.
Take the following exchange I had with my roommate last night:
A: ‘It’s raining outside.’
B: ‘Why don’t you order take-out?’
The overt and literal meaning of the statement ‘It’s raining outside’ refers to the weather conditions. Of course, what I meant was: ‘I don’t want to leave the house to do groceries’. Through a series of inferences, my roomie was able to respond to the implicated meaning of the initial utterance.
In conversation design, we want to assume conversational implicature — your users will use different, less overt and less literal ways to convey their messages. However, in our bot’s responses, we want to focus more on explicature, which is the opposite of conversational implicature.
Conversation design techniques like implicit and explicit confirmations help us with that. A confirmation establishes if the inferred meaning was understood correctly. Additionally, we have to write clear prompts to make sure the user will not infer a different, less overt meaning to what we say.
Information structure
Information structure is a term used to describe how language can be used by speakers to allow them to modify the way information is presented. A widely recognized strategy people use during conversation is to start with what they already know.
The following principles of ‘end focus’, ‘front focus’ and ‘end weight’ all build off of this premise.
End focus principle
Research has shown that people have unconscious expectations about how information is laid out. Therefore, sentences should be presented in a manner that makes it easy for the hearer to decode in time. One approach to the premise of starting with established information is to think about the contrast between information that is ‘given’ and information that is ‘new’.
Generally, we tend to process texts most efficiently when given information is presented before new information. It means that if you want the reader or hearer to take in new information, a good place for it to be is at the end of your sentence. This is called the end focus principle.
As a conversation designer, you should understand that end focus violations can cause unnecessary friction in your interaction with your users. The fact that it is a universal or near-universal principle of language provides a good reason for embracing its functionality.
Front focus
Doing the unexpected is a great strategy for drawing attention to something. So, if you really want to grab your hearer’s or reader’s attention, it is also possible to deliberately put new information at the front of the sentence.
By putting new information where it is not expected, you signal to your audience that you have something to highlight.
End weight
In addition, we also understand sequences in which short and simple sequences come before long and complex ones better than the other way around. As a result, people tend to put these ‘heavier’ elements towards the end of the sentence. This is the principle of ‘end weight’.
As with front focus, there is a range of strategies to ensure that end weight is managed appropriately. Consider the following example:
A: ‘Chatbots, if they communicate in a way that is natural and humanlike, would probably be used by most people on a regular basis.’
The writer of this sentence has failed to follow the end weight principle. As a result, it comes off sounding rather verbose. When we follow the principle of end weight, we could edit as follows:
A: ‘Most people would probably use chatbots on a regular basis if they communicate in a way that is natural and humanlike.’
Conversation analysis
Conversation analysis aims to describe, analyze and understand talk as a basic and constitutive feature of human social life.
Sociologist Harvey Sacks (1935–1975) is generally credited with founding the discipline. He came up with the concept of an ‘adjacency pair’, the smallest unit of conversational exchange: what is in the first part of the pair determines what needs to be in the second part. In other words, it is one utterance with one matching response.
Other topics that have been researched and written extensively about are ‘turn-taking’ and ‘conversational repair’. These two are particularly important to the conversation designer.
Turn-taking
Turn-taking describes the manner in which orderly conversation normally takes place. The word itself already gives you the basic idea: people take turns when speaking to each other.
Understanding the concept of turn-taking is crucial when it comes to voice design. People give each other cues to let each other know when it’s their turn to speak. Turning to someone. Gesturing, a nod. However, in voice design visual cues are excluded from the conversation. And when we rely solely on voice, it is harder for someone to know when it’s their time to speak.
Ending your bot's responses with a clear prompt is an effective strategy. We can also guide the user’s focus through word order, stress, and vocal pitch.
Turn-taking is considered the gearbox of conversation design. By actively moving the conversation forward, keeping responses short and simple, and asking questions your bot can become a true conversational partner.
Conversational repairs
Conversations are messy. People misunderstand each other all the time. Luckily, we are also experts at getting our conversations back on track.
In conversations between people, four types of repair can be identified:
- Self-initiated self-repair, where repair is both initiated and carried out by the speaker of the trouble source.
- Other-initiated self-repair, where repair is carried out by the speaker of the trouble source but initiated by the recipient.
- Self-initiated other-repair, where the speaker of a trouble source may try and get the recipient to repair the trouble.
- Other-initiated other-repair, where the recipient of a trouble source turns both initiates and carries out the repair.
When your voice assistant asks a user to repeat himself or to clarify their intent, that would be a classical example of other-initiated self-repair. Because it is the assistant that asks the user to ‘repair’ the conversation and get back on track.
Another scenario would be when a voice assistant does an implicit confirmation and a user does a barge-in, saying something like ‘that’s not what I meant’ or repeating what they actually meant. This would be an example of self-initiated self-repair, where the user tries to ‘repair’ proactively because he is the one that recognizes there is a misunderstanding.
Why should I care?
So why should you care who repairs the conversation and when to do it?
Because studying how people repair conversations show us that — even though conversations are messy and unpredictable — we still follow certain rules. And rules provide us with clarity.
In the end, knowing the rules will help us design conversations that match people’s conversational strategies. It means we need to learn bots how to speak like people — and not the other way around. Only then we will succeed in creating natural and intuitive user experiences.
Even in the case of misunderstanding.
A career in conversational AI
In conclusion, anyone who’s studying linguistics can have a bright career in conversational AI.
They are needed in a quickly growing market, where the main goal is to advance trust and communication between humans and AI.
Already, lots of computational linguists work for AI companies developing NLU engines and advanced language models to push natural language understanding. But equally important is natural language creation.
Because this is where the knowledge of linguistics is of tremendous value. It means you understand the complexities of how people process speech, handling misunderstandings, and everything else that is inherent to human communication.
Leveraging this knowledge will move the whole field forwards.
