NLP (Natural language processing) has taken great strides recently however how much does Artificial Intelligence understand of what it reads? Short of what we thought, according to scientists at USC’s Department of Computer Science. In an ongoing paper Assistant Professor Xiang Ren and Ph.D. student Yuchen Lin found that despite advances, Artificial Intelligence still doesn’t have the common sense needed to create conceivable sentences.
Lin says, “Current machine next-generation models can write an article that may be convincing to many humans, but they are basically mimicking what they have seen in the training phase.”
“Our goal in this paper is to study the problem of whether current state-of-the-art text-generation models can write sentences to describe natural scenarios in our everyday lives.”
Understanding scenarios in real life
In particular, Lin and Ren tried the models’ capacity to reason and appeared there is a huge gap between the current text generation models and human execution. Given a bunch of verbs and common nouns, cutting edge NLP computer models were tasked with making conceivable sentences describing an ordinary situation. While the models created grammatically right sentences, they were regularly logically incoherent.
For example, here’s one model sentence created by a cutting edge model utilizing the words “dog, frisbee, throw, catch”:
“Two dogs are throwing frisbees at each other.”
The test is based on the assumption that coherent ideas (in this case: “a person throws a frisbee and a dog catches it,”) can’t be generated without a deeper awareness of common-sense concepts. In other words, common sense is more than just the correct understanding of the language it means you do not have to explain everything in a conversation. This is a fundamental challenge in the goal of developing generalizable Artificial Intelligence however beyond academia, it is relevant for consumers, too.
Without an understanding of language, chatbots, and voice assistants built on these state-of-the-art natural-language models are vulnerable to failure. It is also crucial if robots are to become more present in human environments. After all, if you ask a robot for hot milk, you expect it to know you want a cup of a mile, not the whole carton.
“We also show that if a generation model performs better on our test, it can also benefit other applications that need commonsense reasoning, such as robotic learning,” said Lin. “Robots need to understand natural scenarios in our daily life before they make reasonable actions to interact with people.”
Joining Lin and Ren on the paper are USC’s Wangchunshu Zhou, Ming Shen, Pei Zhou; Chandra Bhagavatula from the Allen Institute of Artificial Intelligence; and Yejin Choi from the Allen Institute of Artificial Intelligence and Paul G. Allen School of Computer Science & Engineering, University of Washington.
The common-sense test
Common-sense reasoning, or the ability to make inferences using basic knowledge about the world–like the fact that dogs cannot throw frisbees at each other has resisted Artificial Intelligence scientist’s efforts for decades. State-of-the-art deep-learning models can now reach around 90 percent accuracy, so it would seem that NLP has gotten closer to its goal.
However, Ren, an expert in natural language processing, and Lin, his student, needed more convincing about this statistic’s accuracy. In their paper, published in the Findings of Empirical Methods in Natural Language Processing (EMNLP) conference on Nov. 16, they challenge the effectiveness of the benchmark and, therefore, the level of progress the field has actually made.
“Humans acquire the ability to compose sentences by learning to understand and use common concepts that they recognize in their surrounding environment,” said Lin.
“Acquiring this ability is regarded as a major milestone in human development. But we wanted to test if machines can really acquire such generative commonsense reasoning ability.”
To evaluate different machine models, the pair developed a constrained text generation task called CommonGen, which can be used as a benchmark to test the generative common sense of machines. The researchers presented a dataset consisting of 35,141 concepts associated with 77,449 sentences. They found the even best performing model only achieved an accuracy rate of 31.6 percent versus 63.5 percent for humans.
“We were surprised that the models cannot recall the simple commonsense knowledge that ‘a human throwing a frisbee’ should be much more reasonable than a dog doing it,” said Lin. “We find even the strongest model, called the T5, after training with a large dataset, can still make silly mistakes.”
It seems, said the researchers, that previous tests have not sufficiently challenged the models on their common sense abilities, instead of mimicking what they have seen in the training phase.
“Previous studies have primarily focused on discriminative common sense,” said Ren.
“They test machines with multi-choice questions, where the search space for the machine is small usually four or five candidates.”
For instance, a typical setting for discriminative common-sense testing is a multiple-choice question answering task, for example: “Where do adults use glue sticks?” A: classroom B: office C: a desk drawer.
The answer here, of course, is “B: office.” Even computers can figure this out without much trouble. In contrast, a generative setting is more open-ended, such as the CommonGen task, where a model is asked to generate a natural sentence from given concepts.
“With extensive model training, it is very easy to have a good performance on those tasks. Unlike those discriminative commonsense reasoning tasks, our proposed test focuses on the generative aspect of machine common sense” Ren explains.
Ren and Lin hope the data set will serve as a new benchmark to benefit future research about introducing common sense to natural language generation. In fact, they even have a leaderboard depicting scores achieved by the various popular models to help other researchers determine their viability for future projects.
“Robots need to understand natural scenarios in our daily life before they make reasonable actions to interact with people,” said Lin.
“By introducing common sense and other domain-specific knowledge to machines, I believe that one day we can see AI agents such as Samantha in the movie Her that generate natural responses and interact with our lives.”