AI software may be able to generate text that is grammatically correct and very human-like, but when it comes to common sense they’re still lagging severely behind us humans.
A team of computer scientists from the University of Southern California (USC), the University of Washington, and the Allen Institute for Artificial Intelligence, all in the US, devised a new test examining verbal reasoning skills in machine learning systems. Given a list of simple nouns and verbs, the natural language processing models were tasked with stringing together a sentence to describe a common scenario.
For example, the words “dog”, “frisbee”, “throw”, “catch” prompted one model to generate the sentence: “Two dogs are throwing frisbees at each other.” Although the text is coherent, it’s not something that humans would come up with. The idea of canines playing a game of frisbee isn’t too outlandish, but it’s more plausible that it’d be a human throwing an object for a dog to catch.
“In fact, in our paper, the AI models’ generation is also mostly correct grammatically,” Yuchen Lin, a PhD student at USC, told The Register.
“Their problem is low plausibility - AI generations are either very unusual or impossible in everyday life. For example, “a trash bin is under or on the table” are both grammatically correct but ‘under’ is better for common sense.“
Researchers made an OpenAI GPT-3 medical chatbot as an experiment. It told a mock patient to kill themselvesREAD MORE
The researchers built a dataset made up of 35,141 scenarios described using 77,449 sentences generated by humans, and have tested eight different language models so far. The best performing one known as KG-BART, developed by academics at the University of Chicago, had an accuracy rate of 32.7 per cent compared to Google’s T5-Base model at 22 per cent, according to the leaderboard. All machine learning systems, however, scored lower than humans, who were generally accurate 63.5 per cent of the time.
“For evaluating a model for our proposed task, we use several popular automatic metrics for machine generation: BLEU, METEOR, CiDER, and SPICE. These metrics are basically programs that can give a score between model generations and human references that we collect from many people,” Lin explained.
“BLEU and METEOR are more designed for tasks that machine translation which have a focus on exact word match. Rather, CiDER and SPICE are designed for storytelling, and thus are more suitable for our tasks because we are also open to different scenarios.”
Lin and his colleagues suggest that if AI models don’t have common sense, applications like voice-activated assistants or robots will be prone to mistakes when interacting with humans. Neural networks often fail to develop reasoning skills because they rely on memorizing their training datasets and don’t have a real-world understanding.
“Current machine text-generation models can write an article that may be convincing to many humans, but they’re basically mimicking what they have seen in the training phase,” said Lin.
He hopes that by developing the common sense test, researchers will be able to build better algorithms in the future. “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,” he concluded. ®