Since the beginning of AI, it has been understood that these systems need to learn common sense. In addition, it is now clear, after many years of work, that developing common sense reasoning systems is a labour- and resource-intensive endeavour. 

All humans share a certain amount of common sense. Such information is implicit and undocumented; we know it. We start to pick it up the moment we're born subconsciously. "Animals don't play guitars" or "My grandma is older than me" are examples. Expert humans frequently resort to this body of information when tackling problems with particular domains.

Inferences

A commonsense reasoning process can be supported by commonsense information, leading to inferences like "You might cook food because you want people to eat the food." When coupled with a natural language processing system, the commonsense knowledge base can attempt to respond to questions about the world. Amid the uncertainty, common sense knowledge can be invaluable. 

AI systems

AI systems, like people, use commonly held views about daily items or common sense information to create common sense assumptions or default assumptions about the unknown. It is written as "Normally P holds," "usually P," or "typically P, so Assume P" in an AI system or standard English. If we know that "Tweety is a bird," and we also know that "typically birds fly," then we may reasonably conclude that "Tweety can fly," even if we don't know anything else about Tweety. 

The AI system can adjust its view of Tweety through "truth maintenance" as new information is discovered or acquired. For truth maintenance, let's say we learn that "Tweety is a penguin." This assumption must be rethought since penguins aren't known for their flying abilities.

Commonsense reasoning

Humans can develop assumptions about the nature of the regular circumstances they meet daily based on their experience and training, and they can adjust their "minds" if new information contradicts their previous assumptions. Consider the elements of time, causality, and informational gaps. An essential feature of explainable AI is the capacity to articulate causal relationships. Because they keep detailed logs of assumptions, truth maintenance algorithms inherently furnish an explanation facility. 

Benchmark tests

The Winograd Schema Challenge and other contemporary "commonsense reasoning" benchmark tests show that no extant computer programme comes close to human performance. Some argue that compassionate intelligence is also required for human-level AI. Thus the problem of achieving human-level competency at "commonsense knowledge" tasks is probably "AI-complete" (that is, solving it would require the ability to synthesise a fully human-level intelligence). Common sense reasoning has proven effective in specific fields like natural language processing and computerised diagnosis or analysis.

Commonsense assertions

One of the longest-standing problems in artificial intelligence is creating exhaustive databases of commonsense assertions (CSKBs). Significant improvements were made from early expert-driven initiatives like CYC and WordNet by the crowdsourced OpenMind Commonsense project, which led to the crowdsourced ConceptNet KB. Many methods have been proposed to automate the creation of CSKBs; text mining (WebChild, Quasimodo, TransOMCS, Ascent) and harvesting these straight from pre-trained language models (AutoTOMIC) are just a few examples. 

These databases are far more extensive than ConceptNet, but because of their automated production, they are often of substantially inferior quality. However, there are problems with how common sense is represented; for example, most CSKB projects use a triple data format that could be more optimal for deconstructing more advanced claims made in natural language. GenericsKB stands apart since it doesn't normalise sentences in any way and instead keeps them intact.

Existing applications

BullySpace, an extension of the commonsense knowledge base ConceptNet, was developed by MIT researchers in 2013 to monitor social media for bullying comments. To help the system determine that statements like "Put on a wig and lipstick and be who you are '' are more likely to be insulting if directed at a boy than a girl, BullySpace added over 200 stereotype-based semantic assumptions.

AI systems that write their own stories and chatbots have used ConceptNet. To detect breaches of a complete nuclear test prohibition pact, researchers at Lawrence Livermore National Laboratory developed an intelligent software agent that uses common sense information.

Image source: Unsplash