Extraction of Food Product and Shop Names from Blog Articles Using Named Entity Recognition

Abstract

Named Entity Recognition (NER) is an important research topic in natural language processing, and has been widely studied for a long time. Recently, NER models using neural networks have been proposed, and have achieved the high performance for formal texts written in English such as the dataset of CoNLL2003. We are currently developing a system that provides useful information about food souvenirs that can be purchased only at the particular location or area (called the local limited food souvenirs). Therefore, it is important to apply an efficient NER method to extract food souvenirs and shop names as named entities from noisy user-generated texts written in Japanese such as texts of blog, Q&A, and online review sites. However, most of the existing NER methods using neural networks have not been evaluated with noisy user-generated texts or texts in languages other than English. In this paper, we propose a Conditional Random Fields-based model and compare the performance of the proposed CRF model and the existing state-of-the-art models using neural networks on a dataset constructed by blog texts written in Japanese. From the experimental results, it turned out that a simple neural network model with part-of-speech embeddings for NER has shown the best performance for named entities that were not included in the training data, and the proposed CRF model has shown the best precision in our task.