Feature Subset Selection Using Genetic Algorithm for Named Entity Recognition

Feature Subset Selection Using Genetic Algorithm for Named Entity Recognition Md. Hasanuzzaman1, Sriparna Saha2 and Asif Ekbal2 1 West Bengal Industrial Development Corporation, Kolkata, India Email: hasanuzzaman.im@gmail.com 2 Heidelberg University, 69120 Heidelberg, Germany Email:sriparna.saha@gmail.com, asif.ekbal@gmail.com ⋆⋆ Abstract. In this paper, genetic algorithm (GA) is utilized to search for the appropriate fea- ture combination for constructing a maximum entropy (ME) based classifier for named entity recognition (NER). Features are encoded in the chromosomes. The ME classifier is evaluated for the 3-fold cross validation with the features, encoded in a particular chromosome, and its average F-measure value is used as the fitness value of the corresponding chromosome. The proposed technique is evaluated for determining the suitable feature combinations for NER in three resource-constrained languages, namely Bengali, Hindi and Telugu. Evaluation re- sults show the effectiveness of the proposed approach with the overall recall, precision and F-measure values of 71.27%, 83.95% and 77.09%, respectively for Bengali, 74.72%, 87.15% and 80.46%, respectively for Hindi and 60.91%, 94.15% and 73.97%, respectively for Telugu. Keywords: Genetic algorithm, Feature Selection, Maximum Entropy, Named Entity Recog- nition. 1 Introduction Named Entity Recognition (NER) is a well-established task that has immense importance in many Natural Language Processing (NLP) application areas such as Information Retrieval, Information Extraction, Machine Translation, Question Answering and Automatic Summarization (Babych and Hartley, 2003; Nobata et al., 2002) etc. The objective of NER is to identify and classify every word/term in a document into some predefined categories like person name, location name, organization name, miscellaneous name (date, time, percentage and monetary expressions etc.) and “none-of-the-above”. The main approaches to NER can be grouped into three main categories, namely rule-based, machine learning based and hybrid approach. Rule based approaches focus on extracting names using a number of handcrafted rules that yield better results for restricted domains; and are capa- ble of detecting complex entities that are difficult with learning models. These types of systems are often domain dependent, language specific and do not necessarily adapt well to new domains and languages. Nowadays, researchers are popularly using machine learning approaches for NER because these are easily trainable, adaptable to different domains and languages as well as their maintenance are also less expensive. The main shortcoming of machine learning approach (par- ticularly, supervised systems) is the requirement of large annotated corpus in order to achieve reasonable performance. Thus, building NER systems using machine learning approaches for the resource constrained languages is a great problem. In hybrid systems, the goal is to combine rule- based and machine learning based techniques, and develop new methods using strongest points from each one. Although, hybrid approaches can attain better result than some other approaches, but the weakness of rule-based system still exists when there is a need to change the domain and/or language of data. ⋆⋆ All authors have equal contributions PACLIC 24 Proceedings 153 In the literature, a lot of works are available that use any of these techniques. But, the languages covered include English, most of the European languages and some …

Content truncated to 500 words