Building a true AI assistant starts with processing language
Natural Language Processing is one of AI's hottest fields, yielding both practical products in the marketplace and bleeding-edge research.
The big picture: A virtual assistant that you can truly converse with — which depends on highly accurate speech and text recognition — is still beyond the horizon, but the field is making real progress.
What's happening: Ground is being gained so quickly in NLP that technical advances are threatening to outpace the benchmarks used to test for them, according to the new AI Index report.
- The SuperGLUE benchmark for language understanding tasks was launched in 2019 to replace the existing GLUE standard, which had to be updated because AI models kept exceeding it.
- There was initially a nearly 20-point gap between the best AI systems and human performance, but by January, systems from Microsoft and Google had surpassed humans on SuperGLUE, which asks models to carry out tasks involving answering questions, language inferences and making sense of word disambiguation.
What they're saying: "If we can build an AI that can read PDF files, Word files, websites and the like, then we can actually build something that knows how to answer almost anything we need to deal with," says Igor Jablokov, CEO of the NLP startup Pryon.
Of note: Raleigh-based Pryon last month launched its first commercial product — a virtual assistant platform that can process large amounts of data and use it to answer user questions — for banks and tech companies.
The bottom line: An AI that can read is just as important as one that can write.
Go deeper: AI is industrializing