Conversation Design is an extremely fascinating field that offers businesses greater insight into their sales journey and user experience. Being highly experienced, Master Of Code Global develops AI solutions to help you provide personalized, proactive services. Learn about Conversation Design, which makes your chatbot a modern digital assistant with many advantages.
What is Human-Centric Conversation Design?
Chatbots have come very far, and even that is an understatement. Unfortunately, most consumers do not see the extent of technological advancement barring a few exceptions such as Amazon Alexa, Google Assistant, Microsoft Cortana, and Apple Siri.
This is because most commercial chatbots do not use human-centric conversation design. Most do not try to “chat” with the consumer but rather speak in the same way a sales representative would while reading off a script. While research shows that understanding customers’ needs – and exceeding their expectations – are becoming table stakes for businesses to compete.
Are scripted chatbots bad? No. But they can be better.
And that’s exactly the issue we want to address in this article – how to design a chatbot that is objectively better, not just at achieving business goals but also at solving problems.
Conversation Design in Chatbots
There is a growing disconnect between what chatbots do and what people want chatbots to do. This happens when the technology powering these chatbots gets outdated. New technologies are exciting and intriguing but over time, they lose their spark and consumers become uninterested. This phenomenon isn’t new, nor is it limited to chatbot UX design.
Even with the relatively short history of consuming-facing chatbots, we’ve already been at similar crossroads multiple times before and the only way to overcome this hurdle is to innovate. But innovate how?
Traditionally, the main objective of chatbots is to achieve specific business goals, such as:
But human-centric conversation design, on the other hand, focuses more on personal, human objectives, such as:
We do this because solving human problems is an extremely powerful motivator that drives consumers further into the sales funnel. This means a consumer-first approach is likely to generate greater revenue than a business-first approach would. But there are challenges to this as well.
Companies like IBM, Apple, Google, have been following a similar design approach for a long-time. But because these applications used to require immense computing power, teams of psychology and speech experts, custom AI speech engines, and more, they never fully translated to the enterprise level in terms of capabilities.
But now, all of these barriers have been lowered because smaller enterprises have the tools and technologies to effectively create chatbots similar to Alexa or Siri.
Conversatin Design: Balancing Psychology and Technology
We know a lot about speech and about the human psyche when it comes to speaking and reading. But the challenge comes with translating all of this knowledge into code. It’s so difficult that most developers and companies do not pursue human-centric conversation design at all. Fortunately, there are tools and technologies available today that can be leveraged to not have to start from scratch.
A prime example of such a technology is Google Dialogflow – a cloud-based chatbot platform/builder. There are many other platforms similar to Dialogflow but we want to talk about Dialogflow for two reasons:
- Google is extremely invested in conversational AI. They have their own language model (BERT), their own hardware dedicated to ML-intensive workloads (TPUs), and their Natural Language Understanding (NLU) engine is one of the most well-developed AI speech engines out there.
- The second reason is that conversation design isn’t all about speech – a chatbot also requires a visual interface and Dialogflow has one of the best visual interfaces for creating these chatbots.
Still, the complexity of human-centric conversation design can push developers to opt for scripted chatbot design, which, even with thousands of custom triggers and responses, is still “simpler” than building a human-centric conversational chatbot.
But that’s changing. Dialogflow’s Knowledge Base for instance allows developers to just include resources (articles, books, guides, etc) which the chatbot engine can use to find and relay relevant information to the consumer.
Conversation Authoring within the Intent-Entity-Context-Response (IECR) Paradigm
Designers have two main approaches at their disposal for applying the practice of conversation authoring to chatbots: information-retrieval and natural-language-generation.
The information-retrieval approach involves using a repository of existing queries and responses. The chatbot uses a set of logical conditions to match the user’s query with an appropriate response. These conditions are part of what’s called the IECR paradigm.
Intent-Entity-Context-Response (IECR) Paradigm
All Intent-Entity-Context-Response (IECR) systems are finite-state machines that have mean tree data structure at different levels. Each conversation begins at the root level and is then checked against a series of hierarchical nodes that represent our logical conditions (Intents-Entities-Contexts) until one of the responses matches the query.
Recognizing the intentions of the consumer is paramount in being able to drive the conversation, essentially, keeping it alive. We code possible intents into the framework during chatbot UX design as natural language classes. The chatbot compares the query with all of the predefined intents and uses natural language classification to recognize similar strings of words. If the query matches with a class, the chatbot returns a response from the same natural language class.
If it does not, the chatbot goes to the next level in tree data structure – entities.
Entities are keywords that a chatbot looks for as a source of additional information about the conversation. The chatbot uses exact string matching to find entities, for example, a flight assistant chatbot would identify the keyword “NYC” (and its synonyms) to show all flights to New York City.
For a holistic and efficient approach, chatbots will use exact string matching (entities) and natural language classification (intents) together.
Human speech is very complex which means that many times the same sentence has two, distinct meanings under two scenarios. When creating the chatbot framework, we recognize the common contexts under which a conversation could be initiated and we train the algorithm to use their understanding of contexts to more effectively match responses with triggers.
Take a look at the chatbot exchange below. The chatbot is able to interpret and respond accurately, even when the user answers ambiguously “I don’t know, surprise me”.
By teaching a chatbot to interpret contexts, it is no longer entirely dependent on the messages and thus can deliver the right answer even when the user isn’t providing clear instructions – very realistic. Furthermore, contextual chatbots will also consider previous user preferences to under the conversation better – as seen in the last message.
While very effective, information-retrieval is a limited ability to create new responses. As a result, the natural-language-generation approach, which can synthesize new responses using an AI speech engine is becoming more popular and is often used for conversation authoring in commercial conversational agents. Unfortunately, these chatbots face limitations as well, the most concerning of which is unintended bias.
Benefits of Conversation Design
Lower agent costs
Increased user engagement
Biases in Conversation Design
Chatbots can be trained with massive datasets that may contain a lot more information than what the target audience requires. In such cases, bias may be programmed by the developers to favor certain topics and responses more than others.
But the bias of AI chatbots can also develop unintended bias, which is usually a consequence of complete training or skewed datasets. One of the most popular instances of an AI chatbot using datasets incorrectly is Microsoft’s Tay which learned multiple (toxic) biases from Twitter within 24 hours. Other cases include chatbots using Wikipedia pages and comments to form different biases.
Similarly, another study found that Google Home has a serious accent bias as users with a non-native accent are 30% more likely to get inaccurate responses. One of the reasons for this was that Google Home’s AI engine wasn’t trained with far fewer non-native accents (compared to native English accents).
Similarly, a UNESCO study revealed that the majority of voice-based conversational agents today have female avatars with female voices, indicating a serious gender bias. And like other unintended biases, this has serious consequences. For one, female chatbots tend to promote the stereotypical role of women as assistants. Such stereotypes can be harmful, especially for children and young adults.
Unintended, representational biases take different forms, but often disadvantage or discriminate against certain races, social classes, religions, age, body size, height, etc. But despite being a complex problem, conversational AI can be eliminated. Human-centric chatbots are specifically designed to solve problems and they do this by identifying a larger range of contexts, intents, and human cues. This also helps chatbots steer the conversation and actively avoid confrontations or unfriendly exchanges that are possible due to incomplete data training.
Additionally, we also take conscious steps to combat this unconscious problem, such as:
- Transparency with users
As businesses being transparent with your customers about the capabilities and limitations of chatbots can make users far more forgiving. Being open about AI can also lead to users helping report unintended bias from the chatbot with the intention to help improve it.
- Ethical training
Ethical training here refers to deliberately training the chatbot with a diverse set of datasets to ensure that the chatbot responds in a similar way to all of its users, irrespective of their personal attributes. Google using more non-native speakers to train is Home device would be an example of ethical training.
- Targeted bias
Sometimes it is important to program bias into the chatbot to reduce unintended bias. The premise behind this is that by creating use-case specific chatbot, we can create conversations that are always contextually relevant. For instance, ensuring our banking chatbot prefers talking about banks and not politics. Similarly, a chatbot that is biased to its own bank’s services is likely to result in even fewer issues for the company.
- Testing specifically for unintended biases
Testing for unintended biases by humans is an integral part of our chatbot design. By putting the chatbot in various scenarios, we can evaluate its ability to respond to specific groups of users and ensure it does not pick up on any unintended biases or stereotypes, especially if the chatbot is designed to do a wide variety of tasks (i.e, not use-case specific).
Increase Conversions with Careful Сonversation Design
When deciding whether or not to deploy chatbots, the ROI is one of the biggest concerns – will the chatbot help increase conversions? The answer has and always will depend on how you plan the conversation with your users. Conversation design is complex and doing this five years ago would’ve been much more difficult but today, we have identified the components that help us make the conversation more human. With a human-centric design, we make it easier to communicate with a chatbot and more importantly get things done. This results in significantly higher conversion rates.
To increase conversions with conversational chatbots, it is important to take into consideration your user’s personal objectives, preferences, and even pet peeves when addressing their pain points. In the end, it all comes down to understanding consumer behaviorisms and consumer psychology.
In many ways, human-centric conversation design is just a tried and tested method of developing a personalized and objectively better experience – which ultimately dictates whether or not the chatbot will be successful in maintaining a long-term relationship.
Conclusion: Designing Better Chatbots
Scripted chatbots aren’t inherently worse than chatbots with conversational AI – it’s just that they are very limited in how they “talk”. For instance, they do not accept typed responses which in itself is very unhuman-like. Going forward, our goal should be to use human-centric conversation design and the first step in doing that is to adopt conversational AI which opens up a whole new world of possibilities.
Ready to take your customer experience to the next level?
Talk to an expert and see how you can get started.
Please help us to process your request by answering 3 quick questions
Thank you for your reply!