The market for conversational agents is exploding. Three years ago, there were no fewer than 300,000 conversational agents on Facebook. A study by Comm100the Canadian provider of customer service and communication products, reveals that its conversational agents alone manage more than 25,513 conversations per month. Their customers come from a variety of sectors: banking, education, healthcare and consumer products.
The Office québécois de la langue française defines a conversational agent as a virtual assistant integrated as third-party software into an instant messaging service that can converse with the consumer using natural language, or perform various actions commanded by the consumer.
Typical interaction between a conversational agent and a customer
Hello, how can I help you?
Hi. I’d like to know more about conversational agents.
We currently have 71 articles on conversational agents. What exactly would you like to know?
I’m curious about the use of conversational agents in business. I’d like to know who uses conversational agents the most.
Would you like to know the statistics of conversational agents?
Yes, that’s it. I want to see numbers.
One moment please…
Conversational agents have taken customer service by storm. They alone generated revenues of around US$40.9 million in 2018, and that’s just the beginning. For service sectors such as banking, experts predict that 90% of customer interaction will be automated by 2022. Gartner estimates that by next year, 72% of customer interactions will involve emerging technologies such as machine learning applications, conversational agents and mobile messaging. Finally, Statista predicts that the global market for conversational agents will reach US$454.8 million by 2027.
Contrary to popular belief, consumers appreciate conversational agents; more than half the users surveyed are satisfied, and around 60% of millennials are already using them to buy basic goods.
The limits of written surveys
Would conversational agents be just as effective for conducting written surveys? This was the question on the minds of the management team at INBE, a consumer opinion and survey company based in Quebec City. The company specializes in testing new products for its clients, most of which are SMEs: it sends the target consumer product samples, asks them to test them and answer a few questions.
Like any marketing research professional, INBE has to deal with the limitations of administered questionnaires, including :
- Rigidity: the questionnaire does not allow for clarification of certain questions;
- The format (closed, dichotomous questions) restricts consumer opinions and assessments;
- When questions are open-ended and the answer given is vague (“it’s okay”), there’s no opportunity to bounce back and ask for explanations;
- In the case of multiple-choice questions, there’s no way of ensuring that the consumer has fully understood the proposed answers;
- Impossible to ask for clarification on answers, or to ensure that the respondent answers all the questions on the form.
CRIM develops a sentiment detection algorithm
INBE therefore called on CRIM and its experts in automatic natural language processing (ANLP) to develop a conversational agent that would conduct a consumer survey for a specific product category – food – to measure one aspect – flavor. Unlike a questionnaire, the conversational agent would leave more room for the consumer’s freedom of expression, allowing them to express their opinions and share their impressions fully.
The algorithm would then automatically detect the sentiment expressed by the respondent, and the conversational agent would continue by dialoguing with the person in order to better understand their appreciation and the characteristics of the product being tested.
To build the conversational agent, CRIM used data from past surveys that included salad dressings, cheeses, seaweed treats, pre-packaged marinated chicken and new fruit juice flavors.
Question 1 from the agent : Hello, how do you like this Dijon mustard?
Answer 1: I didn’t really like it.
– The algorithm detects that the answer is too vague, so the agent asks for clarification.
Question 2 from the agent: What did you dislike about this mustard?
Answer 2: I didn’t like the color.
– We’re not talking about taste, so the agent asks a question about taste.
Agent Question 3: And what did you think of the taste?
Answer 3: I find it too spicy.
To train the algorithm, the CRIM-INBE project team annotated over 14,000 responses to 146 questions from past surveys.
Data used to design the CRIM-INBE conversational agent
Number of questions asked: 146
Number of answers given: 14,261
Number of sentences expressing an opinion on taste: 3,055
Number of sentences expressing a positive/negative feeling: 2,583 / 1,423
This data will be used to train the algorithm to distinguish between the nature of the feelings expressed by consumers, and the appreciation of product attributes such as flavor, texture, price, packaging, appearance, and so on.
At the end of this project, the agent will be able to recognize a sentence expressing taste with 90% accuracy (“I didn’t like it”). It will also be able to detect the appreciation of a product characteristic (“I find it too spicy”) with an accuracy of 73%.
Algorithms and conversational agents combine to market better products
These conversations have the potential to replace the static surveys that INBE currently conducts to obtain consumer opinions. The benefits for INBE’s client companies are clear: by quickly identifying consumer preferences, the conversational agent will reduce the risk of dissatisfaction and offer a product that matches their tastes.
Companies will thus be able to obtain more precise and accurate data on consumer behavior, enabling them to define offers that better match their expectations.
