2005 TowardsLargeScaleOpenDomOntolBasedNEClassif

From GM-RKB
Jump to: navigation, search

Subject Headings: Unsupervised NER Task.

Notes

Cited By

Quotes

Abstract

Named entity recognition and classification research has so far mainly focused on supervised techniques and has typically considered only small sets of classes with regard to which to classify the recognized entities. In this paper we address the classification of named entities with regard to large sets of classes which are specified by a given ontology. Our approach is unsupervised as it relies on no labeled training data and is open-domain as the ontology can simply be exchanged. The approach is based on Harris' distributional hypothesis and, based on the vector-space model, it assigns a named entity to the contextually most similar concept from the ontology. The main contribution of the paper is a systematic analysis of the impact of varying certain parameters on such a context-based approach exploiting similarities in vector space for the disambiguation of named entities.

1. Introduction and Related Work

Named Entity Recognition (NER) systems have typically considered only a limited number of classes. The MUC named entity task (Hirschman & Chinchor 97), for example, distinguishes three classes: PERSON, LOCATION and ORGANIZATION, and the CoNLL1 task adds one more: MISC, while the ACE framework2 adds two more: GPE (geo-political entity) and FACILITY. Further, it has often been shown that it is relatively easy to recognize the PERSON and ORGANIZATION classes due to certain regularities, which renders MUC-like named entity recognition tasks even easier.

[1]

In this paper we propose a more challenging task, i.e. the classification of named entities with regard to a large number of classes specified by an ontology or more specifically by a concept hierarchy. Our approach aims at being open-domain in the sense that the underlying ontology and the corpus can be replaced. In our view this aim can only be accomplished if one resorts to an unsupervised system since providing labeled training data for a few hundred concepts as we consider in our approach is often unfeasible. Some researchers have addressed this challenge and have considered a larger number of classes. (Fleischman & Hovy 02) for example have considered 8 classes: ATHLETE, POLITICIAN/GOVERNMENT, CLERGY, BUSINESSPERSON, ENTERTAINER/ARTIST, LAWYER, DOCTOR/SCIENTIST and POLICE. (Evans 03) considers a totally unsupervised scenario in which the classes themselves are derived from the documents. (Hahn & Schnattinger 98) consider an ontology with 325 concepts and (Alfonseca & Manandhar 02) consider 1200 WordNet synsets. In our approach we consider an ontology consisting of 682 concepts.

Named entity recognition and classification has been so far mainly concerned with supervised techniques, the obvious drawback here being that one has to provide labeled training data for each domain and set of classes (compare (Sekine et al. 98; Borthwick et al. 98; Bikel et al. 99; Zhou & Su 02; G. Pailouras & Spyropoulos 00; Isozaki & Kazawa 02; Chieu & Ng 03; Hendrickx & van denBosch 03)). However, when considering hundreds of concepts as possible tags, a supervised approach requiring thousands of training examples seems quite unfeasible. On the other hand, the use of handcrafted resources such as gazetteers or pattern libraries (compare (Maynard et al. 03)) will also not help as creating and maintaining such resources for hundreds of concepts is equally unfeasible. Interesting and very promising are approaches which operate in a bootstrapping-like fashion, using a set of seeds to derive more training data such as the supervised approach using Hidden Markov Models in (Niu et al. 03) or the unsupervised approach in (Collins & Singer 99).

In this paper we present an unsupervised approach which - as many others - is based on Harris' distributional hypothesis, i.e. that words are semantically similar to the extent to which they share syntactic contexts. There have been many approaches in NLP exploiting this hypothesis, the most influential probably being the work of (Grefenstette 94) on automatic thesaurus construction as well as of (Pereira et al. 93) on building hierarchical clusters of nouns, the work of (Hindle 90) on discovering groups of (semantically) similar nouns as well as the work of (Yarowsky 95) and (Schuetze 98) on Word Sense disambiguation/ discrimination. In particular some researchers have considered using syntactic collocations for named entity recognition (cf. (Cucchiarelli & Velardi 01) and (Lin 98)). More recently, several researchers have addressed the problem of classifying a new term into an existing ontology (Agirre et al. 00; Pekar & Staab 02; Alfonseca & Manandhar 02; Widdows).

In this paper we investigate the impact of using different feature weighting measures and various similarity measures described in (Lee 99). Further, to address data sparseness problems we examine the influence of (i) anaphora resolution in the hope that it will yield more context information as speculated in (Grefenstette 94) (ii) downloading additional textual material from the Web as in (Agirre et al. 00) and making use of the structure of the concept hierarchy or taxonomy in calculating the context vectors for the classes as in (Resnik 93), (Hearst & Schütze 93) or (Pekar & Staab 02). The paper is organized as follows: first, we present our data set in Section 2 and describe our evaluation measures as well as present a few baselines for the task showing its complexity in Section 3. In section 4 we analyze the impact of varying the above mentioned parameters step by step starting with a window-based approach as a baseline. Before concluding we also discuss the results of our approach with respect to other systems performing a similar task.

Footnotes

  1. This is a slighlty modified version of the paper published in the proceedings of RANLP 2005
    1. http://cnts.uia.ac.be/conll2003/ner/
    2. http://www.itl.nist.gov/iaui/894.01/tests/ace/phase1/index.htm

References


,

 AuthorvolumeDate ValuetitletypejournaltitleUrldoinoteyear
2005 TowardsLargeScaleOpenDomOntolBasedNEClassifPhilipp Cimiano
Johanna Völker
Towards Large-scale, Open-domain and Ontology-based Named Entity ClassificationProceedings of the International Conference on Recent Advances in Natural Language Processinghttp://www.aifb.uni-karlsruhe.de/WBS/pci/Publications/ranlp05.pdf2005