Salesforce Research explores natural language processing for summarization, translation, and question-answering, with a focus on making models more interpretable and trustworthy. They discuss the DECA benchmark for multitask learning and using reinforcement learning to align models with human-centered metrics.
Natural language processing is the field of trying to understand natural language in a way that's useful for people. You want to summarize a long list of emails, you want to translate a text from English into French, or you want to answer a question that you have.
We have literally hundreds of languages around the world. Each of them have their own history. Each evolves in a different way. Applying AI to understand and process human language is just a great use case.
I'm interested in translating natural language into executable programs, especially adopting state of the art deep learning approaches. It's a very exciting time given that deep learning is making groundbreaking success at different NLP applications. The idea is to make them more transparent so that when we deploy them in a medical setup or autonomous cars, then we would want humans to trust their judgment.
With multitask learning, it's really interesting to see how little we understand how tasks relate to each other. DECA NLP is what we're calling like the Natural Language decathlon. So it's 10 natural language processing tasks. And this is a suite of tasks that we released about a year ago now as a benchmark for us. Now that we have lots of different tasks in the benchmark, we're able to properly see how we're making progress on that and evaluate multitask learning specifically rather than any one of those tasks individually.
My primary research focus is an area called reinforcement learning. Reinforcement learning focuses on learning in an interactive fashion by using trial and error. One of the ways in which reinforcement learning can really benefit natural language processing is it gives us more flexibility basically to really think about what kinds of metrics people really would care about in terms of language. Many of the challenges in natural language processing can be addressed through reinforcement learning.
The future of AI is technically already here. You can talk to your phone. We have driving assist features. What's exciting is that all of these are seeds. They will grow into much more exciting applications eventually where you can have full blown conversations with an AI. It can teach you all kinds of new, interesting facts that you want to learn about. What we're now seeing is just the beginning of a really exciting path that AI will take.
