# The NLP Day

Yesterday (3/13) was the NLP day, and we have the honor to invite Eduard Hovy to give us three talks about the current status and future of NLP researches. The talks were excellent! The first talk was about how machine learning perform so well then previous AI systems. It seems that computer can be more intelligent not because it is smart, but to be trained trained by huge amount of data. The second talk was “A new semantics: merging propositional and distributional information”. Eduard introduced a new model that combine theories and computability that can be used in machine learning. The third talk was “Text harvesting and ontology constructing using a powerful new method.” Using a simple “$N_p \mbox{ such as } N_p \mbox{ and } N_p *$” query and search on Google seems to be trivial, but the results are awesomely incredible! Thus we can mark the “is-a” relationship automatically and the data is satisfying and robust, plus, much more than any relationships in existing wordnets! There are so many interesting topics to write down, but for mow, I’ll focus on the second talk first. You can find the slide for the second talk here.

Though machine learning told us that “you don’t have to be smart, you just need enough training data,” deep in our heart, we still believe that all of our human behaviors is not a huge training table. There must be some rules that guide us, as a theory, to know what we will do and what wouldn’t. We believe that there are theories that can measure info contents, not only treat input strings as meaningless characters that the only purpose is to be sent to machine to do statistical analysis.

### Defining a concept in a new way

A concept $C$ is a list of triples

For example, a dog can be represented as 1

And it can be expressed in a more complexed form:

which we can visualize it as below:

Furthermore, since every node itself can be another context, we can generate a graph! 2

### The main questions to address

• Construction
• Do we really need tensors? Or are vectors enough?
• How to build them?
• Which relations to use? Which scores?
• How do we evaluate this?
• Compositionally: How to ‘add’ tensors to obtain new, more complex, meanings that are still tensors?
• Dependency: How to represent and manage the underlying interconnections across tensor elements?
• Logical operators: How to handle negation, quantification models (can, must…), etc.?

1. I use the syntax of common lisp to express the relationship

2. The copyright of the images belongs to Eduard Hovy