Heuristic Methods for Reducing Errors of Geographic Named Entities Learned by Bootstrapping

Heuristic Methods for Reducing Errors of Geographic Named Entities Learned by Bootstrapping Seungwoo Lee and Gary Geunbae Lee Department of Computer Science and Engineering, Pohang University of Science and Technology, San 31, Hyoja-dong, Nam-gu, Pohang, 790-784, Republic of Korea Abstract. One of issues in the bootstrapping for named entity recogni- tion is how to control annotation errors introduced at every iteration. In this paper, we present several heuristics for reducing such errors using external resources such as WordNet, encyclopedia and Web documents. The bootstrapping is applied for identifying and classifying fine-grained geographic named entities, which are useful for applications such as in- formation extraction and question answering, as well as standard named entities such as PERSON and ORGANIZATION. The experiments show the usefulness of the suggested heuristics and the learning curve evalu- ated at each bootstrapping loop. When our approach was applied to a newspaper corpus, it could achieve 87 F1 value, which is quite promising for the fine-grained named entity recognition task. 1 Introduction A bootstrapping process for named entity recognition is usually as follows. In the initial stage, it selects seeds and annotates a raw corpus using the seeds. From the annotation, internal and contextual patterns are learned and applied to the corpus again to obtain new candidates of each type. Several methods are adopted to reduce over-generation and incorrect annotation and accept only correct ones. One sense per discourse heuristic may also be adopted to expand the annotated instances. It repeats until no more new patterns and entities are learned. There are several issues in bootstrapping approaches for named entity recog- nition task to achieve successful performance. One of them is how to control annotation errors introduced in the bootstrapping process, on which we are fo- cusing in this paper. As iteration continues, the bootstrapping expands previous annotation to increase recall. But this expansion may also introduce annotation errors and, as a result, decrease the precision. Ambiguous entities may be mis- classified since learning speed per class depends on seeds. For example, ‘New York’ may be misclassified to a city name before the patterns that correctly classify it to a state name are learned. Especially such errors in the early stage of the bootstrapping are quite harmful because the errors are accumulated. The R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 658–669, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 Heuristic Methods for Reducing Errors of Geographic Named Entities 659 annotation errors are classified into following four cases: inclusion, crossing, type conflict and spurious. The first three errors occur when a learned entity overlaps a true entity whereas the last one occurs when a learned entity does not overlap any true entity. Most previous works depend only on the statistics (e.g., scores of patterns) obtained from the previous annotation to control such errors. How- ever, this strategy is not always the best because some trivial errors can also be corrected by simple heuristics. We suggest several heuristics that control the annotation errors in Section 4. The …

Content truncated to 500 words