Feature-Rich Named Entity Recognition for Bulgarian Using Conditional Random Fields

International Conference RANLP 2009 - Borovets, Bulgaria, pages 113–117 Feature-Rich Named Entity Recognition for Bulgarian Using Conditional Random Fields Georgi Georgiev∗ georgi.georgiev@ontotext.com Preslav Nakov†§ nakov@comp.nus.edu.sg Kuzman Ganchev‡ kuzman@cis.upenn.ed Petya Osenova§ petya@bultreebank.org Kiril Simov§ kivs@bultreebank.org Abstract The paper presents a feature-rich approach to the automatic recognition and categorization of named entities (persons, organizations, loca- tions, and miscellaneous) in news text for Bulgar- ian. We combine well-established features used for other languages with language-specific lexi- cal, syntactic and morphological information. In particular, we make use of the rich tagset annota- tion of the BulTreeBank (680 morpho-syntactic tags), from which we derive suitable task-specific tagsets (local and nonlocal). We further add domain-specific gazetteers and additional unla- beled data, achieving F1=89.4%, which is com- parable to the state-of-the-art results for English. Keywords Named entity recognition, information extraction, conditional random fields, linear models, machine learning, morphology. 1 Introduction The earliest work on named entity recognition (ner) was based on hand-crafted rules using pattern match- ing [1]. For instance, a rule could encode the knowl- edge that a sequence of capitalized words ending in ‘Inc.’ is typically the name of an organization. An ex- ample of such a system is annie in the gate architec- ture [9]. Such systems could achieve very high preci- sion, but typically suffered from low recall. They also required significant manual tuning, which was time- consuming and could be quite complicated when thou- sands of rules interact in complex manners. Since the nineties, statistical models have offered a viable alternative while requiring little or no manual tuning at all. Such models typically treat ner as a sequence tagging problem, where each word is tagged with its entity type if it is part of an entity. Generative models such as Hidden Markov Models (hmm) [3, 26] have demonstrated excellent performance on the Mes- sage Understanding Conference (muc) datasets [6]. ∗Ontotext AD, 135 Tsarigradsko Ch., Sofia 1784, Bulgaria †Department of Computer Science, National University of Singapore, 13 Computing Drive, Singapore 117417 ‡Department of Computer and Information Science, Univer- sity of Pennsylvania, Philadelphia, PA, USA §Linguistic Modelling Laboratory, Institute for Parallel Pro- cessing, Bulgarian Academy of Sciences, 25A Acad. G. Bonchev St., 1113 Sofia, Bulgaria Discriminative models such as locally-normalized maximum-entropy [4] and conditional random fields (crf) [15] have also been explored for ner. Collins [7] used an hmm tagger to generate n-best outputs, which he reranked discriminatively. By using a semi-Markov crf, [22] recast ner as a segmentation rather than a tagging problem, thus allowing for richer feature sets. Recent research also includes semi-supervised meth- ods, e.g., [16] use word clusters derived from large sets of unlabeled data in order to enrich their feature set. ner can also be viewed as a two-stage process: (1) find the named entities in a sentence, and (2) classify each entity into its type, e.g, person, organization, lo- cation, etc. [7] mentions that first identifying named entities without classifying them alleviates some data sparsity issues. [8] focus on the second stage, named entity classification, assuming that the …

Content truncated to 500 words