In 2012, artificial intelligence researchers revealed a big improvement in computers’ ability to recognize images by feeding a neural network millions of labeled images from a database called ImageNet. It ushered in an exciting phase for computer vision, as it became clear that a model trained using ImageNet could help tackle all sorts of image-recognition problems. Six years later, that’s helped pave the way for self-driving cars to navigate city streets and Facebook to automatically tag people in your photos.
In other arenas of AI research, like understanding language, similar models have proved elusive. But recent research from fast.ai, OpenAI, and the Allen Institute for AI suggests a potential breakthrough, with more robust language models that can help researchers tackle a range of unsolved problems. Sebastian Ruder, a researcher behind one of the new models, calls it his field’s “ImageNet moment.”
The improvements can be dramatic. The most widely tested model, so far, is called Embeddings from Language Models, or ELMo. When it was released by the Allen Institute this spring, ELMo swiftly toppled previous bests on a variety of challenging tasks—like reading comprehension, where an AI answers SAT-style questions about a passage, and sentiment analysis. In a field where progress tends to be incremental, adding ELMo improved results by as much as 25 percent. In June, it was awarded best paper at a major conference.
Dan Klein, a professor of computer science at UC Berkeley, was among the early adopters. He and a student were at work on a constituency parser, a bread-and-butter tool that involves mapping the grammatical structure of a sentence. By adding ELMo, Klein suddenly had the best system in the world, the most accurate by a surprisingly wide margin. “If you’d asked me a few years ago if it was possible to hit a level that high, I wouldn’t have been sure,” he says.
Models like ELMo address a core issue for AI-wielding linguists: lack of labeled data. In order to train a neural network to make decisions, many language problems require data that’s been meticulously labeled by hand. But producing that data takes time and money, and even a lot of it can’t capture the unpredictable ways that we speak and write. For languages other than English, researchers often don’t have enough labeled data to accomplish even basic tasks.
“We’re never going to be able to get enough labeled data,” says Matthew Peters, a research scientist at the Allen Institute who led the ELMo team. “We really need to develop models that take messy, unlabeled data and learn as much from it as possible.”
Luckily, thanks to the internet, researchers have plenty of messy data from sources like Wikipedia, books, and social media. The strategy is to feed those words to a neural network and allow it to discern patterns on its own, a so-called “unsupervised” approach. The hope is that those patterns will capture some general aspects of language—a sense of what words are, perhaps, or the basic contours of grammar. As with a model trained using ImageNet, such a language model could then be fine-tuned to master more specific tasks—like summarizing a scientific article, classifying an email as spam, or even generating a satisfying end to a short story.
That basic intuition isn’t new. In recent years, researchers have delved into unlabeled data using a technique called word embeddings, which maps how words relate to each other based on how they appear in large amounts of text. The new models aim to go deeper than that, capturing information that scales up from words up to higher-level concepts of language. Ruder, who has written about the potential for those deeper models to be useful for a variety of language problems, hopes they will become a simple replacement for word embeddings.
ELMo, for example, improves on word embeddings by incorporating more context, looking at language on a scale of sentences rather than words. That extra context makes the model good at parsing the difference between, say, “May” the month and “may” the verb, but also means it learns about syntax. ELMo gets an additional boost by gaining an understanding of subunits of words, like prefixes and suffixes. Feed a neural network a billion words, as Peters’ team did, and this approach turns out to be quite effective.
It’s still unclear what the model actually learns in the process of analyzing all those words. Because of the opaque ways in which deep neural networks work, it’s a tricky question to answer. Researchers still have only a hazy understanding of why image-recognition systems work so well. In a new paper to appear at a conference in October, Peters took an empirical approach, experimenting with ELMo in various software designs and across different linguistic tasks. “We found that these models learn fundamental properties of language,” Peters says. But he cautions other researchers will need to test ELMo to determine just how robust the model is across different tasks, and also what hidden surprises it may contain.
One risk: encoding biases from the data used to train them, so doctors are labeled as men, and nurses as women, for example, as word embeddings have previously done. And while the initial results generated by tapping ELMo and other models are exciting, says Klein, it’s unclear how far the results can be pushed, perhaps by using more data to train the models, or by adding constraints that force the neural network to learn more effectively. In the long run, AI that reads and talks as fluently as we do may require a new approach entirely.