A Hybrid Learning for Named Entity Recognition Systems

Abstract

This paper presents a hybrid method using machine learning approach for Named Entity Recognition (NER). A system built based on this method is able to achieve reasonable performance with minimal training data and gazetteers. The hybrid machine learning approach differs from previous machine learning-based systems in that it uses Maximum Entropy Model (MEM) and Hidden Markov Model (HMM) successively. We report on the performance of our proposed NER system using British National Corpus (BNC). In the recognition process, we first use MEM to identify the named entities in the corpus by imposing some temporary tagging as references. The MEM walkthrough can be regarded as a training process for HMM, as we then use HMM for the final tagging. We show that with enough training data and appropriate error correction mechanism, this approach can achieve higher precision and recall than using a single statistical model. We conclude with our experimental results that indicate the flexibility of our system in different domains.