Named Entity Learning and Verification: Expectation Maximization in Large Corpora

Named Entity Learning and Verification: Expectation Maximization in Large Corpora Uwe QUASTHOFF, Christian BIEMANN, Christian WOLFF CS Institute, Leipzig University Augustusplatz 10/11 Leipzig, Germany, 04109 Abstract The regularity of named entities is used to learn names and to extract named entities. Having only a few name elements and a set of patterns the al- gorithm learns new names and its elements. A verification step assures quality using a large background corpus. Further improvement is reached through classifying the newly learnt elements on character level. Moreover, unsuper- vised rule learning is discussed. 1 Introduction The task of recognizing person names in text corpora is well examined in the field of Infor- mation Extraction. Most of the approaches, us- ing machine-learning or statistical methods make excessive use of annotated data or large gazetteers. We present a method that needs little input knowledge and performs unsupervised learning on unlabeled data, restricting ourselves to person names. In most languages, named entities form regular patterns. Usually, a surname has a preceding first name, which in turn might have a preceding title or profession. Similar rules hold for differ- ent kinds of named entities consisting of more than one word. Moreover, some elements of such multiwords (like president) are high fre- quency items, well known and of high signifi- cance for identifying a named entity. On the other hand, there are elements that often appear in named entities, but are not characteristic for them (like the first name Israel). Therefore, a verification step is included in the learning algorithm. 2 The Algorithm Our learning algorithm starts with a set of pat- terns and initial name elements. A large corpus of more then 10 million sentences [cf. Quasthoff & Wolff 2000], taken from newspapers of the last 10 years is used for both, the identification of candidates for new name elements as well as for verifying the candidates found. The algo- rithm stops, if no more new name elements are found. The algorithm implements expectation maximi- zation (EM) [cf. Dempster, 1977, Collins, 1999] in the following way: The combination of a learning step and a verification step are iterated. If more name elements are found, the recall of the verification step increases. The key property of this algorithm is to assure high precision and still get massive recall. From another point of view our algorithm im- plements bootstrapping [cf. Riloff 99], as it starts from a small number of seed words and uses knowledge found during the run to find more candidates. 2.1 Patterns and Pattern Rules In a first step the text to be analysed is tagged in the following way: We have two types of tags. The first type is problem dependent. In the case of persons, we have tags for title or profession (TI), first name (FN) and surname (LN). The second tag set is problem independent, but lan- guage dependent. In our experiments, we marked words as lower case (LC) or upper case (UC) depending on the first letter. Punctuation marks are marked …

Content truncated to 500 words