Artificial Intelligence(AI) is simply a process to make machine learn and act smart. McCarthy (2007) describes it as the field of science and engineering of making intelligent machines. It involves designing, understanding and evaluating the machines. These machines can understand natural language, reason various domains of the physical world, and solve new problem by adapting and understanding the problem. British mathematician Alan Turing (1912-54) is regarded as the pioneer in the field of Artificial Intelligence (Berkeley, 1997). In his paper, Turing argued, if a machine could pass his test, we could say the machine was intelligent. This test was later named as ‘Turing Test.’ This was the basic foundation for development of Intelligent Systems. According to Brachman (2002), Intelligent Systems are the systems that know what they are doing. They have the ability to understand, learn and solve new problems.

Although we have a clear definition of what intelligent machines are capable of doing, we have made very little progress in their development. There are multiple factors such as lack of enough research and technology, lack of complete understanding of the physical world, and lack of investment of time and money that have held back the potential in developing AI systems. Brachman (2002) believes other factors such as security, robustness, maintenance, and training come into play to hinder the progress of intelligent machines. According to him, there will be more ways to breach the security if the systems become more complex. More are the chances of systems breaking down when they are made complex. This posts a risk in robustness of the system. It takes more time and effort to maintain large amount of code and requires more training to learn to use the systems. People are worried about robots taking their jobs away. They feel insecure about their job without knowing the fact that more jobs will be created as the technology advances. If we are to dream about a drastic progress in Artificial Intelligence in the next couple of decades, we need to start making people aware about the benefits of advancing technology so that they feel comfortable with the change and join their hands in the development of it.

One of the important fields of Artificial Intelligence is commonsense reasoning and knowledge of the world we live. Current intelligent machines are capable of processing individual words from a text and small domains that perform simple task. However, if it comes to complex systems that require commonsense for human to reason, machines are way behind in the race. There are ambiguous sentences called Winograd Schemas which require the use of commonsense to understand the context of the words in the sentence. It is not possible for a machine to interpret different meaning of the same word in two different sentences without inferring the context of the word with nearby words. This level of inference is very far from what we have achieved in artificial intelligence right now. Techniques such as taxonomic, probabilistic and temporal reasoning have already been implemented. They have tried to lower the gap between human level of inference and computer reasoning. Davis (2011) provides examples of Winograd Schemas. According to him, Winograd schema is a pair of sentences that differ in few words creating ambiguity that can be resolved using real world knowledge and reasoning. One of the examples of Winograd Schemas from his list is, “The trophy doesn’t fit into the brown suitcase because it’s too [small/large]." The ‘it’ in this sentence refers to two different things depending on the last word of the sentence. If the last word is small, the ‘it’ pronoun refers to the suitcase, whereas if the last word is large, the ‘it’ refers to the trophy. For a literate human, there is apparently no ambiguity in the sentence. Human can tell without any hesitation to what the ‘it’ refers. But, a computer has to have a deeper understanding of the context of the entire sentence to get it right. In such cases, the ability of commonsense reasoning is required for the computer. Davis and Marcus (2015) describe about knowledge most important for a computer to understand ambiguous sentence such as Winograd schemas. According to them, computers would require the knowledge of characteristics relations between words and actual understanding of real-world knowledge. In the above example, a machine would require the knowledge of characteristics relation of trophy and suitcase with the fitness; no linguistic inference can help achieve this.

A similar problem occurs when machines try to process natural language. Due to insufficient knowledge of real-world, AI have not yet achieved the perfect smoothness in natural language processing. A difficulty that may arise during language processing is not having related words nearby (Davis and Marcus, 2015). Google translate is a good example to explain this scenario. It is found that Google uses corpus of text to match the nearby words to tackle ambiguity. If two words appear together in most of the texts, the chance that they will have the same meaning in other context is high. Web mining and statistics shows that Google uses this idea for its natural language processing tool. The problem with this idea is whenever those two words are not placed nearby in a sentence. In such case, it becomes difficult for a computer to decide the right meaning. Other difficulties in processing natural language include absence of deep understanding of the text and relevant domain knowledge, and understanding sophisticated and highly tuned rules for matching words (Davis and Marcus, 2015). A large number of web documents have wide range of text rules and characteristics that are very difficult for a simple AI computer to comprehend.

There are other domains in which intelligent systems have been successful in representing and reasoning. Davis and Marcus (2015) state five different AI implemented fields- taxonomic reasoning, temporal reasoning, action and change reasoning, event calculus, and qualitative reasoning. Data structure concepts such as inheritance, transitivity and description logic are used to infer commonsense in taxonomic reasoning. Likewise, temporal reasoning involves reasoning about time duration and time interval. Such reasoning concepts are implemented using system of linear inequalities and mathematical logic, but inference is difficult in this technique too. Action and change reasoning has been implemented using finite state automata and is especially used in high level robotic planning (Davis and Marcus, 2015). According to Davis and Marcus, event planning can be used to represent narratives and temporal reasoning. Integrating different action description of a process is a difficult job because it requires an AI machine to make a decision regarding the depth of abstraction. The authors state qualitative reasoning as the one technique that has seen particular success in commonsense reasoning. This type of reasoning method identifies the direction of change in interrelated qualities. The logic involves one thing going up when other goes up and vice versa.

Studies have been conducted to analyze the difficulties in the automation of commonsense knowledge and reasoning. The need of rich understanding of the world and formulation of a general purpose solution is the single most important difficulty in commonsense reasoning in AI. No single software or device with the strategy for inferring knowledge and making creative decision has been developed till now. The reason behind it being the absence of sufficient technology to handle unexpected events that may arise in a domain. Today’s machines lack the understanding of how different actions of objects interrelate with each other. Instead of deterministic reasoning, many AI applications use probabilistic reasoning which provides a range of answers. This in turn will hold the ability of AI application to determine the depth of abstraction of a problem domain. Money comes as an important factor to consider while developing intelligent systems. Most of the software projects are result oriented which is secure for stakeholders to invest their money on the project. AI reasoning, on the other hand, can be sometimes frustrating and time consuming. We cannot really be sure about the usability of an AI system practically as much as we can be theoretically. AI domain such a human-computer interaction, robotics and natural language processing require complex interpersonal interaction including emotions such as fear, anxiety and happiness. Till date, no ways have been found to represent such states. Robots have been made very similar to human in terms of appearance, but no way near in terms of brain.

Dealing with commonsense reasoning is a challenging task for computer scientist all around the world. However, they have come up with multiple approaches for tackling it. Web Mining is a widely used tool to extract commonsense knowledge from web documents. It is used to infer simple knowledge from multiple documents. Using search pattern, Web Mining has compiled a taxonomy of 2.8 million concepts and 12 million isA relations, with 92% accuracy (Davis and Marcus, 2015). It is amazing that the pattern of words found around the web can help build intelligent programs without the knowledge of syntax and semantics. However, with the benefits of web mining comes its side effects too. The data collected from web mining can be inconsistent and misleading, and is still insufficient to achieve a good level of commonsense reasoning.

There is a high prospect for future development in Artificial Intelligence. The tremendous increase in the processing capability and speed of new processors has had big impact in the software development arena too. Among all the sciences, Computer Science is a new science and Artificial Intelligence is the latest. Within a span of few decades, AI machines has grown from nothing to beating human in games like Chess and Jeopardy. Robots are serving food in restaurants in Japan and automatic satellites revolving around the outer space are sending data back to earth. AI is still in its beginning field and requires a lot of exploration to be done. If more research, time and money is applied in Artificial Intelligence, it has the potential to change the face of our present world.

References

McCarthy, J. (2007, November 12). Basics Questions. Retrieved April 6, 2016, from http://www-formal.stanford.edu/jmc/whatisai/node1.html

Berkeley, I. (1997). What is Artificial Intelligence? Retrieved April 6, 2016, from http://www.ucs.louisiana.edu/~isb9112/dept/phil341/wisai/WhatisAI.html

Brachman, R. J. (2002, November). Systems That Know What They’re Doing. Retrieved April 6, 2016, from http://web.stanford.edu/class/cs227/Readings/Systems That Know What They Are Doing.pdf

Davis, E. (2011, September 8). A Collection of Winograd Schemas. Retrieved April 6, 2016, from https://www.cs.nyu.edu/davise/papers/WS.html

Davis, E., & Marcus, G. (2015, September). Commonsense Reasoning and Commonsense Knowledge in Artificial Intelligence [Scholarly project]. In COMMUNICATIONS OF THE ACM. Retrieved April 6, 2016.