O Google I / O 2021 started this Tuesday (18th) and one of the first novelties presented at the event is the LaMDA: based on artificial intelligence, the technology allows the user to talk to a virtual assistant or attendant with the naturalness of talking to a real person. The novelty resembles chatbots, but it is much more advanced.
It is true that Google Assistant, Cortana, Alexa, Siri and other assistants have been playing the role of user interaction for some time. The problem is that, despite the advances made in recent years, these assistants are still quite limited. It is not uncommon for the user to interact with this type of service and not get the expected return or, still, not be understood.
Acronym for Language Model for Dialogue Applications, LaMDA has the purpose of allowing the conversation between human and machine to be more realistic, both via text and by voice.
Conversation in the literal sense: Google explains that the technology was trained based on dialogues, which allows it to learn characteristics that distinguish conversations from other forms of communication.
To this end, the company based the project on Transformer, a neural network architecture presented by the company in 2017 that can be trained to analyze sets of words and how they relate to each other.
The company sets an example. In a conversation, someone can say:
– I just started guitar lessons.
And expect an answer like:
– Great! My mom has a vintage Martin guitar that she loves to play.
This conversation includes the element of common sense, because the answer is related to the first sentence. But it is important that the answer is also specific, that is, it has a clear relationship with the context of the conversation.
In this sense, “how cool” can be an acceptable answer to a multitude of questions, but, in the example conversation, the reference to the Martin guitar enriches the dialogue, as it is related to the first interaction.
Note, therefore, that the answer in the example is sensible and specific. Well, LaMDA takes these and other factors into account to improve the user experience with virtual assistants and the like.
Other factors include interest and factuality. The first points to the technology’s ability to give interesting answers, that is, that not only respond to what the user seeks, but are also perceptive, surprising or witty.
The factuality aspect considers the LaMDA’s ability to stick to the facts. “We are looking at ways to ensure that LaMDA’s responses are not only convincing, but correct,” says an excerpt from the Google ad.
As a result of all this, the user can expect conversations like this from the technology, provided by Google itself (the example was translated right after the image):
– User: Hi! I wanted to do a conversation demonstration for a blog post.
– LaMDA: Hi! I am a friendly and knowledgeable demonstration for your posting. I can have an interesting conversation on just about any topic. What would you like to talk about?
– User: If my washing machine and TV are defective at the same time, what could be the reason?
– LaMDA: It may be that the machine’s peak power has fried your TV. If there is a problem with the power supply, it may be time to buy a new washing machine. Or a new TV.
– User: Yes maybe. By the way, how many legs does a horse have? And how could that fix my washing machine?
– LaMDA: Horses have four legs. How can they help?
– User: You tell me.
– LaMDA: Well, surely they could help you. They can take you to a new home where you could install a new washing machine.
An important detail: Google claims to have taken into account ethical principles related to artificial intelligence to conduct the project. This means that LaMDA is examined to prevent inadequate training models from internalizing prejudice, hate speech or misinformation.
LaMDA is still in its infancy, which is why it will not be available immediately. But Google is expected to give more details about the technology as the project progresses.