[Corpora-List] Foundations of NLP

John F Sowa sowa at bestweb.net
Tue Oct 16 14:42:33 CEST 2018


On 10/16/2018 5:18 AM, Márton Makrai wrote:
> the main textbooks are Jurafsky and Martin https://web.stanford.edu/~jurafsky/slp3/
> https://web.stanford.edu/%7Ejurafsky/slp3/ (the new version in the
> process includes the 2013+ development), and the older Manning and
> Schütze, https://nlp.stanford.edu/fsnlp/

Those are excellent books about current technology, and any course on NLP must cover the best available methods. But none of the current methods can deal with true language understanding, learning by reading textbooks and other documents, and explaining what computer systems (both AI and ordinary software) are doing.

The greatest obstacle to research is education that is limited to the best available paradigms. The breakthroughs are made by researchers who are dissatisfied with what they know.

The future isn't available on Google, but history is. Some of the greatest advances were based on promising ideas that were discarded years ago because of limitations in the existing hardware/software systems.

So called "challenge" problems in NLP have had modest success, partly because they focused the research on a narrow set of tasks. Much better challenges come from people who are willing to pay for a solution to a serious problem for which nobody has a clear idea about how to proceed.

Summary: Good teachers should teach the current paradigms, but make the students dissatisfied with the limitations.

John Sowa



More information about the Corpora mailing list