Google can always be described as a prominent name for helping organize and access information transmitted using written and oral formats. The company has now announced that it has created machine learning techniques to help it understand the intent of search queries.
Google said during its I/O event that language is one of the most difficult puzzles in computer science, and an even more challenging piece of it is conversations. As a stepping stone to solving this puzzle, the search giant introduced LaMDA, the Language Model for Dialogy Applications. This technology offers a free-flowing capability on a seemingly endless number of topics, more natural ways to interact with technology, and uncovering entirely new categories with a variety of potential applications.
Google explains in its blog post that “while conversations tend to revolve around certain topics, their open-ended nature means they can start from one point and end in a completely different place. A conversation that starts with a friend about a TV show is about the country the show was filmed in. Once you create a discussion, it can turn into a discussion about the country’s best regional cuisine.”
LaMDA’s speech capabilities are built on the Transformer, which also forms the basis for many new language models such as BERT and GPT-3. Transformer is an open source neural network architecture created by Google Research in 2017.
This architecture produces a model that can be trained to read many words, notice how those words are associated, and predict which words it thinks will come next. Unlike other languages, LaMDA is trained on dialogue, including the sensitivity of speech context. LaMDA aims not only to provide sensitivity, but also to be specific.
LaMDA is still in development, and sensitivity and specific responses aren’t the only traits Google is trying to embed in its AI model. It aims for answers that are interesting, insightful, persuasive, and fact-based, while at the same time trying to minimize the abuse of technology. Google carefully researches and reviews all information to ensure that responses are non-biased, free of hate speech, and not misleading.
9343
