Özet:
Named Entity Recognition (NER) is the task of detecting and categorizing the entities in a given text. It is an important task in Natural Language Processing (NLP) and forms the basis of many NLP systems. Previous work on NER that make use of statistical models can be categorized into two main categories: feature-based and embedding-based. Earlier work on NER made frequent use of manually crafted fea tures. In order to use manually crafted features we either automatically annotate the dataset for the given features using third party software or manually annotate the dataset, both of which require additional work. Recent work make use of BiLSTM based neural networks and represent words with embeddings. This relieves systems from relying on manually created feature sets. In this work, we started with analyzing the performance of the feature based systems. In this phase, we reimplemented a pre vious work and improved the performance by making use of the dependency parsing features. Following these results, we implemented a novel method that makes use of both dependency parsing features and embeddings. We propose a novel BiLSTM CRF based neural model that makes use of the dependency parsing feature to learn both tasks jointly in a unique way. Our model jointly learns both dependency parsing and named entity recognition using separate datasets for each task. The model does not require the named entity recognition dataset to be annotated for the dependency pars ing task. Our results show that performance increases when we use a joint learning model instead of annotating the named entity recognition dataset automatically.
