At the Allen Institute for AI in Seattle, computer scientist Yejin Choi is leading project Mosaic, which aims to teach machines common-sense knowledge and reasoning, one of the hardest and longest-standing challenges in the field of artificial intelligence (AI).
Choi, senior research manager, leads the project, which started in 2018 and recently delivered its first results. Choi is also an associate professor at the Paul G. Allen School of Computer Science & Engineering at the University of Washington in Seattle.
What is your definition of common sense?
Common sense is about the basic level of practical knowledge and reasoning that concerns everyday situations and events. This is knowledge that is commonly shared among most people. Most 10-year-old kids possess it, but it is very hard for machines. For example: don't leave the door of the fridge open too long, because the food will go bad. Or: if I drop my mug of coffee on the floor, the floor will get wet and the mug might break.
Common sense knowledge is not just about the physical world, but also about the social world. "If Kate smiles, she is probably happy."
Why is it important for machines to have common sense?
Humans use common-sense knowledge and reasoning all the time. Machines need it to understand the needs and actions of humans better, and to reason about their causes and effects, even when faced with novel, uncertain situations. While modern machine learning methods have made dramatic advancements across different AI benchmarks, their high performance is often confined to a particular task and domain. As a result, if presented with examples that are either adversarial or out of domain, they often make surprisingly silly mistakes. These mistakes often have to do with a lack of common sense.
Giving machines common sense is a long-standing problem in the field of AI. Why has there been so little progress for decades?
The best-known large-scale effort to give machines common sense has been Douglas Lenat's CYC project, started in 1984. Lenat's idea was to explicitly program common-sense knowledge in a logical form in a machine. Tens of millions of common-sense statements, like 'water is wet', have been hand-coded in CYC. This approach is expensive, slow, and has turned out to be very brittle.
Why do you think the time is right to finally make progress?
We have given a dramatically different take on the common-sense challenge. Instead of representing knowledge in purely logical forms, we teach neural networks representations of common-sense knowledge and reasoning. This turns turn out to be far more robust.
In the teaching we combine a large-scale symbolic knowledge graph, that corresponds to descriptive knowledge, with a Web-scale corpus of unstructured text, which corresponds to implicit knowledge reflected in natural language. Compared to the approaches from the previous decades, we have modernized common-sense modeling on all fronts: significantly more data, significantly more compute, and new computational methods based on deep learning.
In addition, we employ the full spectrum of natural language as the symbolic representation of knowledge, instead of logical forms. This allows for large-scale crowdsourcing based on Amazon Mechanical Turk, which is much more cost-effective than hand-crafted logical rules written by experts.
How is the Allen Institute for AI trying to give machines common sense?
Our research consists of several research thrusts that range from knowledge graph construction to common-sense reasoning, and from natural language inference to visual reasoning.
One of our research focuses is on building broad-scale common-sense knowledge graphs. Everybody can download and use it. The vast majority of the existing resources largely focused on the knowledge of 'what'; for example, 'water is liquid' and 'coffee is a beverage'. In contrast, our knowledge graph ATOMIC focuses on the knowledge of 'why' and 'how'."
What are some results that you have achieved already?
In ATOMIC, we have built an atlas of common-sense reasoning. Using neural networks, we trained it not only on knowledge graphs, but also on large amounts of text: news articles, blogposts, novels, etc. Combining the two, our model now contains more than one million textual descriptions of 'if-then' relations; for example, 'if X pays Y a compliment, then Y will likely return the compliment'. The model can distinguish causes versus effects, agents versus themes, voluntary versus involuntary events, and actions versus mental states. We have shown that neural models can acquire simple common-sense capabilities and reason about previously unseen events.
In addition to ATOMIC, we also developed COMET, a generative model which can reason about causes and effects of new events that are not yet in ATOMIC. When I type the sentence 'Alex drives his sister to the mall', the system automatically gives possible causes, although it has never seen this sentence before. It suggests that Alex might want to be helpful, might go shopping, or might want to be nice to her. It also reasons about what happened before; for example, that Alex has to start the car. And it can reason about the effects of the action: Alex might feel happy, or he might get thanked by his sister. Our model is quite robust. It also works with longer sentences.
Can you give some examples of how practical AI systems will benefit from having common sense?
Generally, when AI gets more common sense, practical devices will become safer and more useful. For example, speech assistants like Siri or Alexa will understand you much better, and will be able to answer a much wider set of questions.
Future household robots will understand by themselves that some leftover pie first has to be stored in a container of the right size, and then has to be put into the fridge.
Bennie Mols is a science and technology writer based in Amsterdam, the Netherlands.
No entries found