Elon University Home

Elon Computing Sciences

COMPUTERS YOU CAN TALK TO: APPLYING MACHINE LEARNING TECHNIQUES TO NATURAL LANGUAGE UNDERSTANDING

Presentation at Elon Student Undergraduate Research Forum, Spring 2005

Michael M. Richards (Professor Shannon Pollard) Department of Computing Sciences

Artificial Intelligence is still a relatively new area of study within computer science, but has become much more popular as people have started to realize the capabilities and possibilities of computers. As our knowledge of algorithms increases, and the power of the computer tool increases, we have been able to give the computer human-like characteristics.

Machine Learning is an area of Artificial Intelligence in which algorithms are employed to enable a computer to gain knowledge of a particular problem over time. There are several known methods of learning; each method has its own benefits for particular types of problems. The problem we studied is the capability of a computer to grasp meaning from the words of our Natural Language. Some algorithms have been developed for language comprehension, but none have been without fault, which is why any research could be profitable for the area.

To solve the problem of Natural Language comprehension, we utilized several Machine Learning algorithms with the simple game of Tic-Tac-Toe to see how well the computer could eventually understand simple spoken phrases like, “Put an X in the center square.” In the search for the answer to this question, we also had to determine what might be the best method of learning to use.

Supervised reinforcement learning is a technique in which a teacher will give you both an input, or in this case a phrase, and will tell you what the desired output should be. We used this type of learning in the computer and then evaluated the results of the teachings by giving it different phrases to determine if the computer can decipher the right meaning. Through our research, we found that the algorithm for this type of learning worked very well for our problem: our program was able to learn Tic-Tac-Toe dialog after about 20 – 25 interactions, and was able to provide the right output about 80% of the time. It remains to be seen whether or not these techniques can be expanded for use on larger-scale applications. If so, it could one day lead to full Natural Language computer interfaces.