Olga Simek and Courtland VanDam, MIT Lincoln Laboratory, USA
Arabic named entity recognition (NER) is a challenging problem, especially in conversational data such as social media posts. To address this problem, we propose an Arabic weak learner NER model called ANER-HMM, which leverages low quality predictions that provide partial recognition of entities. By combining these predictions, we achieve state of the art NER accuracy for cases for out-of-domain predictions. ANER-HMM leverages a hidden markov model to combine multiple predictions from weak learners and gazetteers. We demonstrate that ANER-HMM outperforms the state-of-the-art Arabic NER methods without requiring any labeled data or training deep learning models which often require large computing resources.
Named entity recognition, Arabic, weak learning.