Chatbots have become more advance and widespread than ever before. From eCommerce to social media, various platforms use the AI conversational agent for a variety of reasons.
As a result, developers have always designed the software to excel at a specific conversational skill or style. The researchers at Facebook AI Research were no different. They also developed chatbots in the same manner.
For example, a few AI models were proficient at talking to humans in a certain way, while another could incorporate knowledge into a conversation. Facebook also had a model that was good at an empathetic response, and another that could talk about its persona consistently.
Humans, on the other hand, are not like that. We can switch to different conversational styles based on the situation that we find ourselves.
For examples, a human speaker can easily alternate between:
- Exchanging information
- Listening and responding empathetically
- Talking about him/herself
Now, the team at Facebook AI Research has created a chatbot that can do all three things.
In a statement to the press, one of the researchers who carried out the study, Eric Smith, said:
“Our goal in this study was to produce a single model capable of smoothly switching between and blending all three of these kinds of communication.”
The researchers explained how they trained the chatbot in their pre-published paper on arXiv.
Developing Chatbots with Multiple Conversational Skills
The Facebook AI team started by training a single model on three sets of conversations.
Along with the different contexts, the researchers designed the three conversations to demonstrate various skills. These are talking about oneself, listening with empathy, and providing knowledge.
“This allowed our model to do well on each of those skill benchmarks in isolation,” Smith said.
The second challenge was teaching the model to switch seamlessly between different conversation skills as humans do.
For this part, the researchers trained the model using a new dataset dubbed BlendedSkillTalk — now available on the ParlAI online platform. The dataset includes roughly 5,000 human dialogue in which speakers switch between the three conversation styles.
Smith said:
“Our study shows how to leverage existing datasets that are focused on one skill each in order to train a model that can blend those skills seamlessly in a conversation.”
The study has some interesting implications for the development of chatbots.
For one, it could lead to more versatile and better performing conversational agents. Other research teams can also use the new dataset to train, evaluate, and compare Natural Language Processing (NLP) models.
Eric Smith and colleagues at Facebook AI Research are now working to extend the conversational skills further.
In addition to the three skills, future chatbots may possess humor and an ability to comment on images, among others.
Comments (0)
Least Recent