i202 Fall 2010 School of Information, UC Berkeley

Published Monday, the NYT wrote about the Never-Ending Language Learning system (NELL). This article touches upon many topics from recent lectures!

Developed by a team of researchers at Carnegie Mellon, NELL is supposed to 'learn' the way humans do. So far, NELL has "scann[ed] hundreds of millions of Web pages for text patterns that it uses to learn facts… These facts are grouped into semantic categories — cities, companies, sports teams, actors, universities, plants and 274 others." NELL was first "primed by the researchers with some basic knowledge in various categories" before she was released into the wild, wild web. NELL is an example of supervised machine learning, where humans had to "train" a system to "understand" certain categories. Once “taught", NELL was then able to produce her own categories and classify information on her own. Impressively, NELL also appears to understand different meanings/ relationships between words, though NELL still manages to misinterpret words (she placed 'internet cookies' under the category 'bake goods’). To further emphasize that Nell is an example of supervised learning, humans regularly check her work and correct her errors. Whenever NELL misinterprets a relationship (internet cookie vs. baked cookie), a human correctly categorizes the word for her, allowing her to learn from her ‘mistakes.’ As a result, NELL is able to "grapple with words in different contexts, by deploying a hierarchy of rules to resolve ambiguity." Thus, the categories allow her to (almost) correctly distinguish and categorize polysemous words. Though used for unsupervised machine language, NELL’s scanning of web pages reminds me of the clustering concept Jess wrote about. It appears that NELL is ‘clustering’ related concepts (keywords, word patterns, words indicating relationships, etc) into defined categories. Thus, can the term clustering be used for supervised machine languages?