Researchers from Facebook introduced Persona-Chat, a database consisting of more than 160,000 fragments of dialogues of real people who spoke on behalf of fictitious persons. The developers also trained a computer model on the example of these data: in the future, as reported in the preprint published on arXiv, it can help personify chat bots and voice assistants.
Despite the fact that voice assistants and chat-bots are getting more and more perfect every year, they can still support the conversation with difficulty. The reason for this is the limitations of the training sample: the computer can be taught to talk and even answer questions, but it may not be enough to maintain a seamless conversation. The chat-bot, for example, does not have a formed personality and interests – and therefore can not answer complex open questions like “what do you like to do on a rainy day?”. Of course, there is an option in which a chat-bot or voice assistant answers an unknown question with an abstract phrase like “I do not know” or a search query – but it can not be considered a conversation. In addition, when communicating with such an interlocutor retains in memory only a small snippet of dialogue and can not return to the information provided to him a few steps back.
Developers from Facebook AI Research, led by Jason Weston, have collected a database of more than a thousand “persons”: small (five sentences of not more than 15 words each) descriptions of abstract personalities invented by real people. For example, a description of such a “persona” might look like this:
“I am a vegetarian. I like swimming. My father worked at Ford. My favorite band is Maroon5. A month ago I started working as a designer in an advertising agency “
In total there are 1155 such “persons” in the database. In addition, the researchers asked another group of people to evaluate the resulting “people” and rewrite them on the basis of similar, related characteristics: for example, the love for Maroon5 could be replaced in the “persona” for the love of the song “She Will Be Loved”, and the statement about that the father of “persona” worked for Ford, could turn into the statement that a close relative worked in an automobile company. This is necessary in order to expand the facts known about the “person” in order to draw up a possible dialogue.After this, two people who provided “persons” were asked to talk a little: each of them was assigned an occasional “persona”. As a result, 164,356 statements were found in the database of dialogues, and on the basis of this database researchers trained several computer models (generative Seq2Seqand ranking Memory Networks ). The resulting models of chat bots were then evaluated by asking them to talk with real people: the dialogue could be built around either the “person” of the computer, or around the “person” of the person, or around both participants in the dialogue. The developed models bypassed the algorithm, trained dialogues on quotes from films, on fluency, involvement in dialogue and completeness of conversation.
An example of a person dialogue (PERSON 1) and a chat-bot trained on the collected database
The developers note that the collected database can be useful for creating new and improving old models of chat bots and voice assistants. The database is also available for public access.Maintaining conversations on abstract topics is more like an addition to voice assistants and chat bots. And their main function, in addition to controlling something and recognizing voice commands, can even be the provision of psychological support: as Woebot, which uses methods of cognitive-behavioral psychotherapy when communicating with users .