Abstract:

A semi-supervised approach based on a three-level framework for product named entity recognition is presented. The structure features and relationships among different parts of product named entities are studied, and a combined method is applied. A hidden conditional random field model is built so as to utilize the hidden status of learned samples. The labels failed to be learned by the bootstrapping algorithm is considered as hidden statuses. Experiment in digital camera area shows that, with only a few manually labeled data, this method could recognize product named entities from text contents of web pages very well.

Key words: product named entity recognition, business information processing, natural language processing