Combining n-gram based statistics with traditional methods for named entity recognition

In this paper, we show three main results. First, we show that an n-gram dataset built from a large web crawl, as opposed to data from the specific target domain, can be used to perform the task of named entity recognition with reasonable accuracy. Second, we show that for complex domains, such as the MUC-7 NER task, the Lex method may not perform as well as other methods, due largely in part to subjective definitions of what constitutes a named entity. Finally, we show that for both clear domains and more subjective domains, combining Lex with traditional lexical approaches can achieve higher accuracy than either alone.