Reading:
Nilsson, N.J. (1980). Principles of Artificial Intelligence
Natural language processing (NLP) has always been a central focus of artificial intelligence development. While humans can effortlessly acquire languages during childhood, it is less explicit how we will be able to encode the complex grammars we employ in our daily communications into a machine and make it capable of extracting and “understanding” the information contained in the sentences. However, even if we can fully understand the grammatical rules behind human languages and encode them into computational forms, I think another layer of difficulty might be encountered before we can build up a machine that is fully competent to communicate with human beings.
One assumption being made behind this approach is that all the information needed to carry out a normal conversation is fully embedded in the sentence itself, which could then be extracted by a machine simply by analyzing its words and syntactic structure. Nevertheless, if we carefully observe human conversations and comprehend them in a more “logical” way, this assumption may not seem that valid anymore. For instance, please take a look at the following dialogue:
Ben: Let’s go to Disneyland tomorrow, Tom!
Tom: Oh, I have a test in the third period.
Ben: All right then. Maybe we can try next Saturday?
If we try to think purely from a computer’s viewpoint, it seems that understanding this dialogue requires more information than the ones contained in the sentences themselves. For example, we need to have the background knowledge of what Disneyland is, what does “go to Disneyland” imply, and how long does it typically take for people to play at Disneyland. In addition, we also need to recognize the relative importance of a test and going to Disneyland, what “the third period” means for Tom (say Tom is an undergraduate student), and how this period of time conflicts with going to Disneyland (as it will typically take a whole day), etc. As we can see, a simple dialogue like this actually involves more background knowledge and common sense about the world than we used to imagine. And for a machine to be able to communicate freely as we do, these shared background knowledge and inferences outside of the codes are essential.
To summarize, I think one of the major difficulties faced by NLP, in addition to the complexities involved in the coding of grammatical structures of human languages, is actually how background knowledge and common sense about the world could be learned by the machines and applied into their processing of sentences. As the information is not fully embedded in the sentences themselves, the real understanding of them will inevitably rely on extracting the “relevant” information based on the “context” of communication, which may also include the “communicative intention” of the other agencies involved in the conversation. As all of these factors are crucial in human communication, their realizability in artificial intelligence is of great importance. Yet, these are extremely difficult problems to deal with as we are even far from having a thorough understanding of our own communication system.