Named Entity Discovery Using Comparable News Articles

Named Entity Discovery Using Comparable News Articles Yusuke SHINYAMA and Satoshi SEKINE Computer Science Department New York University 715, Broadway, 7th Floor New York, NY, 10003 yusuke@cs.nyu.edu, sekine@cs.nyu.edu Abstract In this paper we describe a way to discover Named Entities by using the distribution of words in news articles. Named Entity recog- nition is an important task for today’s natural language applications, but it still suffers from data sparseness. We used an observation that a Named Entity is likely to appear synchronously in several news articles, whereas a common noun is less likely. Exploiting this characteris- tic, we successfully obtained rare Named Enti- ties with 90% accuracy just by comparing time series distributions of a word in two newspa- pers. Although the achieved recall is not suf- ficient yet, we believe that this method can be used to strengthen the lexical knowledge of a Named Entity tagger. 1 Introduction Recently, Named Entity (NE) recognition has been getting more attention as a basic building block for practical natural language applications. A Named Entity tagger identifies proper expressions such as names, locations and dates in sentences. We are trying to extend this to an Extended Named Entity tagger, which additionally identifies some common nouns such as disease names or products. We be- lieve that identifying these names is useful for many applications such as information extraction or ques- tion answering (Sekine et al., 2002). Normally a Named Entity tagger uses lexical or contextual knowledge to spot names which appear in documents. One of the major problem of this task is its data sparseness. Names appear very frequently in regularly updated documents such as news arti- cles or web pages. They are, however, much more varied than common nouns, and changing contin- uously. Since it is hard to construct a set of pre- defined names by hand, usually some corpus based approaches are used for building such taggers. However, as Zipf’s law indicates, most of the names which occupy a large portion of vocabulary are rarely used. So it is hard for Named Entity tagger developers to keep up with a contemporary set of words, even though a large number of docu- ments are provided for learning. There still might be a “wild” noun which doesn’t appear in the cor- pora. Several attempts have been made to tackle this problem by using unsupervised learning tech- niques, which make vast amount of corpora avail- able to use. (Strzalkowski and Wang, 1996) and (Collins and Singer, 1999) tried to obtain either lex- ical or contextual knowledge from a seed given by hand. They trained the two different kind of knowl- edge alternately at each iteration of training. (Yan- garber et al., 2002) tried to discover names with a similar method. However, these methods still suffer in the situation where the number of occurrences of a certain name is rather small. 2 Synchronicity of Names In this paper we propose another method to strengthen the lexical knowledge for Named Entity tagging by using synchronicity of names …

Content truncated to 500 words