Ai Systems Beat Humans at Reading Comprehension

The BERT neural network has led to a revolution in how machines sympathize human language.

In the autumn of 2017, Sam Bowman, a computational linguist at New York Academy, figured that computers yet weren't very good at understanding the written give-and-take. Certain, they had go decent at simulating that understanding in sure narrow domains, like automated translation or sentiment analysis (for example, determining if a judgement sounds "mean or nice," he said). But Bowman wanted measurable evidence of the genuine article: bona fide, human-style reading comprehension in English. And so he came up with a exam.

In an April 2018 newspaper coauthored with collaborators from the University of Washington and DeepMind, the Google-owned artificial intelligence visitor, Bowman introduced a bombardment of ix reading-comprehension tasks for computers called GLUE (General Language Understanding Evaluation). The exam was designed as "a fairly representative sample of what the research community idea were interesting challenges," said Bowman, simply also "pretty straightforward for humans." For case, ane task asks whether a sentence is true based on information offered in a preceding sentence. If you can tell that "President Trump landed in Iraq for the kickoff of a seven-24-hour interval visit" implies that "President Trump is on an overseas visit," yous've just passed.

The machines bombed. Even state-of-the-art neural networks scored no college than 69 out of 100 beyond all ix tasks: a D-plus, in letter grade terms. Bowman and his coauthors weren't surprised. Neural networks — layers of computational connections built in a rough approximation of how neurons communicate within mammalian brains — had shown promise in the field of "natural language processing" (NLP), only the researchers weren't convinced that these systems were learning annihilation substantial about language itself. And Gum seemed to testify it. "These early results indicate that solving GLUE is across the capabilities of current models and methods," Bowman and his coauthors wrote.

Their appraisal would be short-lived. In October of 2018, Google introduced a new method nicknamed BERT (Bidirectional Encoder Representations from Transformers). Information technology produced a Glue score of eighty.five. On this brand-new benchmark designed to measure machines' real agreement of tongue — or to expose their lack thereof — the machines had jumped from a D-plus to a B-minus in just six months.

"That was definitely the 'oh, crap' moment," Bowman recalled, using a more than colorful interjection. "The general reaction in the field was incredulity. BERT was getting numbers on many of the tasks that were shut to what we thought would be the limit of how well you lot could do." Indeed, GLUE didn't even bother to include human baseline scores before BERT; by the time Bowman and one of his Ph.D. students added them to Mucilage in Feb 2019, they lasted just a few months earlier a BERT-based system from Microsoft beat them.

Every bit of this writing, about every position on the GLUE leaderboard is occupied by a system that incorporates, extends or optimizes BERT. V of these systems outrank homo performance.

But is AI actually starting to empathise our linguistic communication — or is it just getting better at gaming our systems? As BERT-based neural networks have taken benchmarks like GLUE past storm, new evaluation methods have emerged that seem to paint these powerful NLP systems as computational versions of Clever Hans, the early on 20th-century equus caballus who seemed smart enough to do arithmetics, but who was actually just post-obit unconscious cues from his trainer.

"Nosotros know nosotros're somewhere in the gray area betwixt solving language in a very boring, narrow sense, and solving AI," Bowman said. "The full general reaction of the field was: Why did this happen? What does this mean? What do nosotros practice at present?"

Writing Their Own Rules

In the famous Chinese Room thought experiment, a not-Chinese-speaking person sits in a room furnished with many rulebooks. Taken together, these rulebooks perfectly specify how to accept any incoming sequence of Chinese symbols and craft an appropriate response. A person outside slips questions written in Chinese under the door. The person inside consults the rulebooks, then sends dorsum perfectly coherent answers in Chinese.

The thought experiment has been used to fence that, no thing how it might appear from the exterior, the person inside the room can't be said to take any true understanding of Chinese. Still, even a simulacrum of understanding has been a good enough goal for natural language processing.

The only problem is that perfect rulebooks don't exist, because tongue is far as well complex and haphazard to be reduced to a rigid set of specifications. Take syntax, for instance: the rules (and rules of thumb) that define how words group into meaningful sentences. The phrase "colorless green ideas sleep furiously" has perfect syntax, but any natural speaker knows it'due south nonsense. What prewritten rulebook could capture this "unwritten" fact about natural language — or innumerable others?

NLP researchers have tried to square this circle by having neural networks write their own makeshift rulebooks, in a procedure called pretraining.

Before 2018, i of NLP'south main pretraining tools was something like a dictionary. Known as word embeddings, this dictionary encoded associations betwixt words as numbers in a way that deep neural networks could accept every bit input — alike to giving the person inside a Chinese room a crude vocabulary book to work with. Just a neural network pretrained with word embeddings is still blind to the meaning of words at the sentence level. "It would call back that 'a human chip the dog' and 'a domestic dog bit the man' are exactly the same thing," said Tal Linzen, a computational linguist at Johns Hopkins University.

A better method would use pretraining to equip the network with richer rulebooks — not merely for vocabulary, but for syntax and context besides — earlier training it to perform a specific NLP task. In early 2018, researchers at OpenAI, the Academy of San Francisco, the Allen Found for Artificial Intelligence and the Academy of Washington simultaneously discovered a clever way to guess this feat. Instead of pretraining simply the outset layer of a network with word embeddings, the researchers began training entire neural networks on a broader basic chore called linguistic communication modeling.

"The simplest kind of language model is: I'thou going to read a bunch of words and and then try to predict the next word," explained Myle Ott, a inquiry scientist at Facebook. "If I say, 'George Bush was born in,' the model at present has to predict the side by side discussion in that sentence."

These deep pretrained linguistic communication models could be produced relatively efficiently. Researchers simply fed their neural networks massive amounts of written text copied from freely available sources similar Wikipedia — billions of words, preformatted into grammatically correct sentences — and permit the networks derive side by side-word predictions on their ain. In essence, it was similar asking the person inside a Chinese room to write all his own rules, using just the incoming Chinese messages for reference.

"The great matter about this approach is it turns out that the model learns a ton of stuff about syntax," Ott said.

What's more, these pretrained neural networks could then apply their richer representations of language to the job of learning an unrelated, more specific NLP chore, a procedure called fine-tuning.

"You can have the model from the pretraining stage and kind of conform it for whatever bodily task you care about," Ott explained. "And when you practice that, y'all become much better results than if yous had just started with your end task in the beginning place."

Indeed, in June of 2018, when OpenAI unveiled a neural network called GPT, which included a language model pretrained on about a billion words (sourced from eleven,038 digital books) for an entire month, its GLUE score of 72.8 immediately took the tiptop spot on the leaderboard. Notwithstanding, Sam Bowman assumed that the field had a long way to go before any system could fifty-fifty brainstorm to approach human-level performance.

And then BERT appeared.

A Powerful Recipe

So what exactly is BERT?

First, it's non a fully trained neural network capable of besting human performance correct out of the box. Instead, said Bowman, BERT is "a very precise recipe for pretraining a neural network." Just as a baker can follow a recipe to reliably produce a delicious prebaked pie chaff — which tin and then be used to make many different kinds of pie, from blueberry to spinach quiche — Google researchers developed BERT's recipe to serve as an ideal foundation for "baking" neural networks (that is, fine-tuning them) to do well on many unlike natural language processing tasks. Google too open-sourced BERT'south code, which ways that other researchers don't have to repeat the recipe from scratch — they can just download BERT equally-is, similar buying a prebaked pie crust from the supermarket.

If BERT is substantially a recipe, what'southward the ingredient list? "Information technology's the issue of three things coming together to actually make things click," said Omer Levy, a research scientist at Facebook who has analyzed BERT's inner workings.

The kickoff is a pretrained language model, those reference books in our Chinese room. The 2d is the ability to figure out which features of a sentence are most of import.

In 2017, an engineer at Google Brain named Jakob Uszkoreit was working on means to accelerate Google's language-understanding efforts. He noticed that state-of-the-fine art neural networks also suffered from a built-in constraint: They all looked through the sequence of words one past 1. This "sequentiality" seemed to match intuitions of how humans actually read written sentences. But Uszkoreit wondered if "it might be the case that understanding linguistic communication in a linear, sequential style is suboptimal," he said.

Uszkoreit and his collaborators devised a new architecture for neural networks focused on "attention," a mechanism that lets each layer of the network assign more weight to some specific features of the input than to others. This new attention-focused architecture, called a transformer, could accept a sentence like "a dog bites the man" as input and encode each word in many different ways in parallel. For instance, a transformer might connect "bites" and "man" together as verb and object, while ignoring "a"; at the same time, it could connect "bites" and "dog" together as verb and subject, while mostly ignoring "the."

The nonsequential nature of the transformer represented sentences in a more expressive form, which Uszkoreit calls treelike. Each layer of the neural network makes multiple, parallel connections between certain words while ignoring others — akin to a student diagramming a sentence in elementary school. These connections are oftentimes drawn between words that may not actually sit next to each other in the sentence. "Those structures effectively look similar a number of copse that are overlaid," Uszkoreit explained.

This treelike representation of sentences gave transformers a powerful manner to model contextual meaning, and also to efficiently larn associations between words that might exist far away from each other in circuitous sentences. "It's a flake counterintuitive," Uszkoreit said, "but information technology is rooted in results from linguistics, which has for a long time looked at treelike models of language."

Finally, the third ingredient in BERT's recipe takes nonlinear reading one footstep further.

Unlike other pretrained language models, many of which are created by having neural networks read terabytes of text from left to correct, BERT's model reads left to correct and right to left at the same time, and learns to predict words in the middle that have been randomly masked from view. For example, BERT might have as input a sentence like "George Bush-league was [……..] in Connecticut in 1946" and predict the masked word in the middle of the sentence (in this case, "born") past parsing the text from both directions. "This bidirectionality is conditioning a neural network to try to become as much data as it can out of any subset of words," Uszkoreit said.

The Mad-Libs-esque pretraining chore that BERT uses — called masked-linguistic communication modeling — isn't new. In fact, it's been used as a tool for assessing language comprehension in humans for decades. For Google, it also offered a practical way of enabling bidirectionality in neural networks, equally opposed to the unidirectional pretraining methods that had previously dominated the field. "Earlier BERT, unidirectional language modeling was the standard, fifty-fifty though information technology is an unnecessarily restrictive constraint," said Kenton Lee, a research scientist at Google.

Each of these iii ingredients — a deep pretrained language model, attention and bidirectionality — existed independently before BERT. Simply until Google released its recipe in late 2018, no 1 had combined them in such a powerful way.

Refining the Recipe

Like whatever good recipe, BERT was soon adapted by cooks to their own tastes. In the spring of 2019, there was a catamenia "when Microsoft and Alibaba were leapfrogging each other week past week, continuing to tune their models and trade places at the number 1 spot on the leaderboard," Bowman recalled. When an improved version of BERT called RoBERTa first came on the scene in August, the DeepMind researcher Sebastian Ruder dryly noted the occasion in his widely read NLP newsletter: "Another month, another state-of-the-art pretrained language model."

BERT's "pie chaff" incorporates a number of structural design decisions that bear on how well information technology works. These include the size of the neural network existence broiled, the corporeality of pretraining data, how that pretraining information is masked and how long the neural network gets to train on it. Subsequent recipes similar RoBERTa result from researchers tweaking these design decisions, much like chefs refining a dish.

In RoBERTa's case, researchers at Facebook and the University of Washington increased some ingredients (more than pretraining data, longer input sequences, more than grooming fourth dimension), took one away (a "next judgement prediction" task, originally included in BERT, that actually degraded performance) and modified some other (they fabricated the masked-language pretraining task harder). The outcome? Start place on GLUE — briefly. Six weeks afterward, researchers from Microsoft and the University of Maryland added their own tweaks to RoBERTa and eked out a new win. As of this writing, yet another model called ALBERT, short for "A Lite BERT," has taken Gum'southward acme spot by farther adjusting BERT'southward bones design.

"We're still figuring out what recipes work and which ones don't," said Facebook's Ott, who worked on RoBERTa.

Still, just as perfecting your pie-blistering technique isn't likely to teach you the principles of chemistry, incrementally optimizing BERT doesn't necessarily impart much theoretical noesis about advancing NLP. "I'll be perfectly honest with y'all: I don't follow these papers, because they are extremely boring to me," said Linzen, the computational linguist from Johns Hopkins. "There is a scientific puzzle in that location," he grants, but it doesn't lie in figuring out how to make BERT and all its spawn smarter, or even in figuring out how they got smart in the commencement identify. Instead, "we are trying to understand to what extent these models are really understanding linguistic communication," he said, and not "picking up weird tricks that happen to work on the data sets that we normally evaluate our models on."

In other words: BERT is doing something right. But what if it's for the wrong reasons?

Clever but Not Smart

In July 2019, ii researchers from Taiwan's National Cheng Kung University used BERT to achieve an impressive result on a relatively obscure natural language understanding criterion called the statement reasoning comprehension task. Performing the task requires selecting the appropriate implicit premise (called a warrant) that will support a reason for arguing some claim. For example, to fence that "smoking causes cancer" (the merits) because "scientific studies have shown a link betwixt smoking and cancer" (the reason), you lot need to presume that "scientific studies are credible" (the warrant), as opposed to "scientific studies are expensive" (which may be true, but makes no sense in the context of the argument). Got all that?

If not, don't worry. Even human being beings don't do particularly well on this task without practice: The average baseline score for an untrained person is 80 out of 100. BERT got 77 — "surprising," in the authors' understated opinion.

But instead of concluding that BERT could evidently imbue neural networks with near-Aristotelian reasoning skills, they suspected a simpler explanation: that BERT was picking upward on superficial patterns in the way the warrants were phrased. Indeed, afterwards re-analyzing their training data, the authors found aplenty evidence of these so-chosen spurious cues. For case, merely choosing a warrant with the word "non" in information technology led to correct answers 61% of the time. After these patterns were scrubbed from the data, BERT'south score dropped from 77 to 53 — equivalent to random guessing. An commodity in The Gradient, a car-learning magazine published out of the Stanford Artificial Intelligence Laboratory, compared BERT to Clever Hans, the horse with the phony powers of arithmetic.

In another paper called "Correct for the Wrong Reasons," Linzen and his coauthors published testify that BERT's loftier performance on sure GLUE tasks might also be attributed to spurious cues in the training information for those tasks. (The paper included an culling data ready designed to specifically expose the kind of shortcut that Linzen suspected BERT was using on GLUE. The data set's name: Heuristic Assay for Natural-Language-Inference Systems, or HANS.)

Then is BERT, and all of its benchmark-busting siblings, essentially a sham? Bowman agrees with Linzen that some of Mucilage'south training data is messy — shot through with subtle biases introduced past the humans who created it, all of which are potentially exploitable by a powerful BERT-based neural network. "In that location'southward no unmarried 'cheap flim-flam' that will let it solve everything [in Glue], only there are lots of shortcuts information technology tin have that will really assistance," Bowman said, "and the model can pick up on those shortcuts." But he doesn't think BERT's foundation is built on sand, either. "Information technology seems like nosotros have a model that has really learned something substantial about language," he said. "But information technology's definitely non understanding English language in a comprehensive and robust way."

According to Yejin Choi, a figurer scientist at the University of Washington and the Allen Institute, one fashion to encourage progress toward robust agreement is to focus not merely on building a improve BERT, but as well on designing ameliorate benchmarks and training data that lower the possibility of Clever Hans–manner cheating. Her piece of work explores an approach called adversarial filtering, which uses algorithms to scan NLP training data sets and remove examples that are overly repetitive or that otherwise introduce spurious cues for a neural network to option up on. Afterward this adversarial filtering, "BERT's performance can reduce significantly," she said, while "human being performance does non drop and so much."

Nonetheless, some NLP researchers believe that even with improve grooming, neural language models may still face a key obstruction to real understanding. Fifty-fifty with its powerful pretraining, BERT is not designed to perfectly model language in full general. Instead, after fine-tuning, it models "a specific NLP job, or fifty-fifty a specific information set up for that task," said Anna Rogers, a computational linguist at the Text Machine Lab at the University of Massachusetts, Lowell. And it'south likely that no preparation data gear up, no thing how comprehensively designed or carefully filtered, tin capture all the edge cases and unforeseen inputs that humans effortlessly cope with when we utilise natural language.

Bowman points out that it's difficult to know how we would ever be fully convinced that a neural network achieves anything like existent understanding. Standardized tests, afterward all, are supposed to reveal something intrinsic and generalizable about the test-taker's knowledge. Just equally anyone who has taken an Sat prep class knows, tests tin can be gamed. "We accept a difficult fourth dimension making tests that are difficult enough and trick-proof plenty that solving [them] really convinces us that we've fully solved some aspect of AI or language technology," he said.

Indeed, Bowman and his collaborators recently introduced a test chosen SuperGLUE that's specifically designed to be hard for BERT-based systems. So far, no neural network tin beat human performance on it. But fifty-fifty if (or when) it happens, does it mean that machines can really understand language whatever better than earlier? Or does just information technology mean that science has gotten better at didactics machines to the test?

"That's a expert analogy," Bowman said. "We figured out how to solve the LSAT and the MCAT, and we might not actually be qualified to be doctors and lawyers." However, he added, this seems to be the manner that artificial intelligence research moves forrad. "Chess felt like a serious test of intelligence until we figured out how to write a chess programme," he said. "We're definitely in an era where the goal is to go on coming upwards with harder problems that represent language agreement, and go on figuring out how to solve those issues."

Clarification: On Oct 17, this article was updated to clarify the signal made past Anna Rogers.

This article was reprinted on Wired.com, in High german at Spektrum.de and in Spanish at Investigacionyciencia.es.

jenkinsworposs.blogspot.com

Source: https://www.quantamagazine.org/machines-beat-humans-on-a-reading-test-but-do-they-understand-20191017/

Share this post