Facebook is reportedly working with researchers from Stanford University to overcome existing limitations in chatbot capabilities.
The number of chatbots used by online businesses is expected to rise in the next two years, but with 65 percent of consumers believing chatbots are “too dumb” to be useful, researchers have set out to improve the level of intelligence and helpfulness of this popular form of AI-tech.
According to a new paper, Learning from Dialogue After Deployment: Feed Yourself, Chatbot, AI research scientists from Facebook and Stanford University believe future iterations of chatbots will include tech that allows the bots to “self-improve”, as it extracts training data from conversations.
In the paper, the authors explained that chatbots will start utilising user responses when conversations are going well, using them as “training examples” that can be imitated in the future. “When the agent believes it has made a mistake, it asks for feedback; learning to predict the feedback that will be given improves the chatbot’s dialogue abilities further.” It’s believed that by only using natural responses from the user for learning, the bot will be less reliant on numerical feedback or human intervention.
However, the researchers do note that by letting the chatbot train itself through real-time conversations there is a possibility it will start to reinforce errors, an issue that AI scientists believe is easy enough to track and rectify, where necessary. For instance, during testing, the researchers contracted workers to have conversations with the bot and to later provide a rating between one to five to determine the quality of each of the bot’s responses. These ratings were then used to teach the bot and help it predict how satisfied or unsatisfied a human user might be based on its responses to questions and its ability to hold a conversation.
The chatbots were also rated on the ability to do two tasks at the same time; what it’s going to say next and the coherency of its replies. Each time the bot interacted with a human, it would take its former exchanges into consideration. If the bot received a low score for its responses, it would then ask for feedback.
According to the researchers from Facebook and Stanford University, chatbot accuracy in terms of responses increased by 31 percent over the course of the study, with the best performing bot achieving 46.3 percent accuracy on dialogue tasks and 68.4 percent on feedback tasks. This represents a much higher success rate than other methods, in a shorter amount of time.
Full details and the training code used for the study will reportedly be made available through Facebook’s ParlAI platform.
What This Means For Retailers
In late 2018, Kylie Gleeson-Long, ANZ managing director of dunhumby said that the convergence of chatbots and virtual assistants “has the potential to transform the retail experience”, as retailers begin to look for new technologies that can enhance customers’ shopping experiences. She also noted that chatbots represent a good data source for online businesses, with bots having a vital role to play in the continued development of the online sector in 2019.
This comes as recent research revealed that 72 percent of consumers feel like chatbots can be useful, allowing the tech used is serviceable. To put this into perspective, only 58 percent of consumers currently rank their experiences with chatbots as “adequate”, while 18 percent believe the AI tech is “ineffective and annoying”.
By 2020, Gartner expects 25 percent of customer service operations will be handled through virtual assistants, the online retail industry included. This represents significant potential for retailers to maximise their customer service potential in a convenient, tech-savvy manner.