Saturday, July 21, 2007

The gist and computational semantics...

Gist is described in Wikipedia as the general significance, of remembered experience. I'm continuing to read up on semantics, computational semantics and how things are interrelated. Language is a way to communicate about events and concepts, but for language to work, we need to have similar ideas about concepts, as otherwise miscommunication occurs. How do we resolve miscommunication and do we adjust our world model to compensate for this? The very interesting part in this is that for communication to work across cultures, we need to understand how other cultures perceive the world and what their norms and values are for communication and interaction.

The learning process is the key to human intelligence. The way I see it now is that we learn a number of concepts through our senses and actions, then as we learn to communicate, we are able to talk about similar things since we have a way to imagine and reconstruct those similarities. Without similarities (learning by analogy?), it'd be very difficult to learn anything. In another line of thought, I can imagine that we start with a large ball of conceptual knowledge that we are born with. Then through continuous experimentation, observation and communication this gets divided into more specialized concepts. Eventually better represented as a certain kind of conceptual network. It is interesting to note that some linguists believe that not all words of our vocabulary are learned from external sources, but may actually be inferred by meaning based on similarity. This is quite difficult to prove however.

Some readings in the area of ontology allowed me to understand a bit more about how we can imagine the storage of concepts in a conceptual tree. When analyzing similarity or thinking about how things are related, the larger the distance between these items, the longer it takes to interpret the concept. This also would explain how it becomes more difficult to learn something if we know nothing about the concept. For example, research indicated that when associating canary with singing, this is easily accepted and very quickly associated and confirmed as true. But the distance between canary and songbird is only single step. Associating a canary with flight takes slightly longer to confirm, as the distance is now two steps (through bird and then to flight). Other relationships may not exist and probably should not be created as they would develop a wrong relationship between concepts and therefore a wrong conclusion or erroneous view of this world. The strength between these relationships (belief?) may be rather difficult to change once it is established.

As we thus grow this conceptual tree and develop relationships and enrich it with splits in certain concepts (differentiation), we constantly re-shape our conceptual network. My doubt in this theory is whether besides real-world concepts that we store in the brain, we also store how we inferred a certain relationship, as this would help us in the future to derive other concepts more rapidly. This helps us to find other similarities at a faster level.

Gist is a project that attempts to analyze concepts in a certain space and makes divisions between these concepts using support vector machine classifications. The research is very interesting and I am wondering whether support vector machines are part of the key to allow machines to learn similarly and build a similar conceptual tree.

I imagine the brain as a very large network. Even though the network cannot yet derive meaning or produce language as we are born, the network or an externality to it must have the ability to train it. Supposing that the chemical and neuro-biological processes in the brain produce a certain sequence or state (that which represents a certain concept), how do we know or test that this state or sequence is that what we observe or listen or is meant by somebody else? This requires us to continuously test these concepts with the external world and re-test our experience against our observations until the network produces something that comes close to the actual experience. This raises the interesting question how we become efficient in testing observation against our idea of the thing, whether they are part of the same network, and so on.

I've ordered a couple of books that allow me to dive into the material for real from an academic perspective. I'm very interested whether the ideas that I have developed are in one way or another similar to existing theories.

No comments: