At Frase, we believe that AI-driven Question Answering is an essential technology capability for future chatbot systems. In this article, we will cover the basics about conversational marketing, the state of AI-driven question answering, and most importantly, why your bot needs additional intelligence to resolve your customer’s questions.
Is Your Chatbot Helping Users Learn From Your Website’s Content?
According to a recent study by SEMRush, the most influential ranking factors Google looks at are related to user engagement: time on site, bounce rate, and pages per session. If you want your content to be on page 1 of Google, you should invest in ways to maximize engagement. This entails the overall customer experience throughout your website.
The question is: when users land in your site through organic search, how do you help them discover and interact with your content?
With that question in mind, we recently surveyed 300 people about their behavior regarding site search (the search engine inside your website, if you have one). Survey data showed that most people would rather leave your website and Google again. Outside of e-commerce where site search is critical for product discovery, it seems like most websites haven’t paid much attention to its importance.
What does this mean to marketers? According to DemandGen Report, 47% of buyers view 3 to 5 pieces of content before engaging with a sales representative. When users land on your website, do you make it easy for them to keep learning? These are some current ways to help users find what they are looking for in your website:
- Site maps: ugly and complicated.
- Navigation menus: limited in design.
- Site search: technologically outdated and worse than Google.
- Chatbots: efficient for marketing, sales, and support, but not great for information retrieval or content discovery.
Let’s focus on chatbots. Unless you’ve lived under a rock for the last couple years, you must have witnessed the explosive growth of website chatbots. When you visit a website, the chances are high that there will be an avatar with a welcome message in the bottom right of the screen.
To the general public, the term “chatbot” implies automation and the idea of holding a conversation with a machine. The reality is that current solutions are far from that paradigm.
The Basics: What is Conversational Marketing?
According to Drift (the company that coined the term), conversational marketing “is a one-to-one approach to marketing that companies use to shorten their sales cycle, learn about their customers, and create a more human buying experience. Instead of forcing people to fill out static lead forms and wait for follow-ups (that might never come), conversational marketing focuses on engaging people in real-time, both with human-to-human conversations and human-to-chatbot conversations.”
How are Chatbots Improving Marketing, Sales, and Customer Support?
- Real-time response: as researched by Drift, ideally you should respond to new leads within five minutes of them reaching out.
- Automated scheduling: chatbots allow your website to engage prospects 24/7. Scheduling a meeting is a relatively simple action that can be automated by a chatbot.
- Lead qualification: chatbots can engage your leads with qualification questions, and integrate customer data with your CRM.
- Increase landing page conversion rates: chatbots can provide a more engaging and personalized experience compared to traditional forms.
- Customer support: chatbots can be integrated with your Help Center to provide quick access to support articles. If your customer still needs extra help, chatbots can route you to the right support agent.
Difference between Chatbots and Question Answering Systems
Now that we know the basics about conversational marketing let’s make sure we draw a line between “Chatbots” and “Question Answering Systems.” These are their basic definitions:
- A chatbot is a computer program designed to simulate a conversation with human users, especially over the Internet (Wikipedia).
- A Question Answering System is a computer science discipline within the fields of information retrieval and natural language processing (NLP), which is concerned with building systems that automatically answer questions posed by humans in a natural language (Wikipedia).
Workflow-based vs. Open Ended
Chatbots are designed to help users complete specific workflows, such as placing an order for a product or signing up for a service. Note that chatbots propose options to the user according to dialogs previously configured by a human (if the user chooses Option A, then reply with Option B, and so forth). Advanced chatbots might let you ask a question freely, but their ultimate goal is to understand your intent and take you down a pre-configured dialog.
In contrast, Question Answering systems are not pre-configured. They are trained to learn from a specific knowledge base (like websites or a document repository) and answer questions flexibly, no rules required.
Machine-Human Hybrid vs. Full Automation
Chatbots are usually a blend of human and machine. Most chatbot solutions provide real-time messaging capabilities to allow a regular conversation between humans. For example, you might see a chatbot greet you with some qualification questions, but you eventually start a conversation with a human.
In contrast, Question Answering systems are meant to be a fully automated experience, more like a search engine specialized to give you fast access to information.
Conversation vs. Informational Retrieval
Chatbots attempt to replicate human conversations. To this end, they usually start by asking qualifying questions to keep the conversation under control. In most cases, chatbots don’t benefit from having you ask open-ended questions.
In contrast, Question Answering systems work oppositely: they excel at taking questions and retrieving precise information from a knowledge base.
Create content with an AI-powered Research Assistant
Limitations of current chatbots
Although there is proof that chatbots have a positive impact on marketing, sales, and customer support, some limitations and challenges are affecting current solutions.
When a website has a chatbot, it is usually pre-configured to accomplish a particular task. The question is: is the chatbot always aware of what the end user cares about? Or is our chatbot mostly focused on helping you accomplish a business goal?
For example, let’s say you run a b2b company selling SEO tools. If someone is reading an excellent blog post you wrote about content marketing, is it relevant for your sales rep to pop up in the screen? Of course, this is something that you can configure in the bot’s settings, but it is always tempting to load the bot everywhere hoping for your readers to engage with your sales team. At the end of the day you are producing content to generate leads, but how do you use your bot without disturbing your readers?
Reliance on hardcoded dialogs
As we discussed, chatbots generally rely on particular dialogs and decision trees. In the example below, the user asked “how does SEO work?”, and the bot attempted to schedule a demo instead of addressing the user’s question – not the most useful user experience.
When chatbots are automated, they rely on a highly structured knowledge base.
Certain chatbots have Question Answering capabilities, especially when it comes down to customer support. Because FAQs are highly structured knowledge bases, it is somewhat “easy” to understand your question and point you to a related support article. The limitation is that this won’t work in situations where your knowledge base is less organized, such as your website, blog or internal documents. Question Answering “in the wild” gets much more complicated.
Question Answering Systems and the Gold Rush of Natural Language Processing
Question Answering has been an active area of academic research for years. With the release of the Squad dataset by Stanford in 2017 (+100k Question-Answer pairs), industry leaders started to build deep learning models that could find answers contained in Wikipedia passages. IBM Watson made headlines with its Jeopardy system, but nobody found production level solutions.
Things changed in 2018 when Google released a new dataset and language model called BERT. This model is meant to provide a universal understanding of language. Traditionally, data scientists had been developing task-specific models (for example, your system for topic extraction would be entirely different than your question answering system), but now you could use BERT to generalize across tasks. It only took a few weeks to see developers leverage BERT to accomplish the state-of-the-art results and achieve human performance in the Squad dataset.
The complexity of end-to-end Open-Domain Question Answering
While progress made on the Squad dataset is excellent, most of the proposed Question Answering systems are not “end-to-end.” They are trained to read a passage and find the correct answer in it. However, there is one big thing missing here: how do you find the right passage when you are searching for thousands or millions of documents?
Retrieving the correct passage across a vast repository of documents has become a new challenge. To this end, Microsoft released a new dataset in 2018 called MS Marco. This dataset includes search results to real life queries made on Bing, along with the correct answer buried in those search results – a much harder task.
This is all high-level information about the state of artificial intelligence for question answering. Comparisons can be drawn with the progress made years ago when Google released breakthrough datasets for computer vision. The conclusions are: we are making major technology milestones, and fast.
What Websites Should Be Leveraging AI-Driven Question Answering?
Not all websites might be able to benefit from a Question Answering bot. However, certain types of sites should consider adding a layer of automated Question Answering:
- Information-heavy websites, like those including regulations, handbooks, FAQs, and technical content that is regularly accessed.
- B2B companies producing recurrent thought leadership content in a given industry.
- Publishers that need to increase page views to increase advertising revenue.
Websites that have already embraced website chatbots might be in a good position to provide Question Answering as an additional capability. For example, companies using Drift or Intercom for conversational marketing might have already experienced the power of bots. In addition, help center tools like Zendesk or Help Scout are already offering chat solutions to manage their knowledge base. These are also good candidates to leverage AI-driven question answering.
Conclusions: Why is 2019 the Year for Question Answering bots?
- Conversational marketing and automated Question Answering systems are complementary and fulfill different tasks.
- Long gone are the days of SmarterChild, and scripted chatbots, technology is coming where you’ll have to think twice, whether or not
theperson helping you is human.
- Think about your chatbot strategy as a multi-layered system: from full automation to human conversations.
- Question Answering remains a significant data science challenge, but we are getting closer to commercialization.
- Investing in tools that increase your website engagement metrics is vital for organic search growth. Both chatbots and Question Answering can contribute to this goal.
- Investing in chatbots and Question Answering means making your content future-proofed for conversational search and voice.