Facebook researchers teaching chatbots how to chat like humans

Facebook seems to be working around the clock to make chatbots enable with the ability to talk. Suggested otherwise by its name, chatbots can't really chat. As explained in a pre-print paper by researchers from Facebook's FAIR lab, chatbots fail to perform this task on multiple levels. They are incapable of maintaining a "consistent personality" and they can't remember what they or their conversational partners have said in the past.

Further, chatbots reply with pre-programmed responses like "I don't know", when they are asked a question they don't understand. Despite will all these shortcomings, chatbots can prove to be engaging. However, we can't have a natural, meaningful conversations with them. In an attempt to make this possible, researchers from Facebook have now turned to deep learning.

This means, instead of feeding them with pre-programmed questions and answers, chatbots will be taught the conversation skills by looking for patterns in large datasets. While there have been impressive developments, the researchers are facing difficulties in choosing the right data to begin with.

For example, some chatbots these days are trained on dialogue taken from movie scripts. Needless to say, you won't be able to have a proper conversation with them as they will keep saying irrelevant lines from the movie script. In order to fix this problem, Facebook's engineers have built their own dataset which the chatbots will learn from.

Called Persona-chat, this dataset comprises of over 160,000 lines of dialogue, sourced from workers found on Amazon's Mechanical Turk marketplace. For those who are unaware, Amazon Mechanical Turk (MTurk) is a marketplace for work that requires human intelligence.

Notably, the Persona-Chat contains is not random. To give chatbots some personality, the Mechanical Turk workers were asked to make the dialogues according to their own characters. They came up with five basic biographical statements to use them inside the conversation.

For example, one persona was based around the following statements: "I am an artist. I have four children. I recently got a cat. I enjoy walking for exercise. I love watching Game of Thrones." No doubt that chatbots have a long way to go, but it is a start.

This data was then fed to the chatbots and the results were later assessed by another group of Mechanical Turkers. Turns out, the persona bot was not as fluent and consistent as the humans, it was much better than the chatbot trained with movie dialogues.

Facebook Messenger Tips and Tricks [MUST WATCH] - GIZBOT

On the flip side, the persona chatbot was not that engaging as it soon ran out of topics to talk about, compared to the chatbot trained with movie dialogues. The researchers from Facebook has not offered any explanations behind this.

Via