Identifying General Product and Brand Names in Online Forums

点击翻译按钮获取中文摘要

One key data component for engaging and acquiring customers is being able to identify names of products inonline forums. Therefore, the extraction of product/brand names will generate data to help vendors and thusincrease overall business revenue. The traditional approach to extract named entities requires time-consumingand expensive manual human-labelled training sets. In addition, we deal with many product or service sectorswhich come with different size, language, and content. Thus fitting a supervised model which is based onhuman-annotated data, for each product/brand to each sector (vertical) can be very expensive. Three majorchallenges are: 1) the high cost of generating training data for each type of product, 2) covering large divergenttypes of products from all verticals, and 3) disambiguating different types of entities based on context.Therefore, in the absence or lack of training data, alternative solution is semi-supervised learning algorithmssuch as bootstrapping and meta-learning methods such as self-training and co-training. In these methods, wecan train a model either with no or very few annotated data. In this research, a hybrid approach combiningtransfer learning and semi-supervised learning is investigated to identify and extract named entities in ourdomain of interest. In transfer learning, the solution for the lack of annotated data in the target domain, is toadapt annotated data from other domains.VerticalScope (the industry partner) is a Canadian company that is becoming a leading player in data scienceresearch and development for understanding user-generated content on the Internet in a variety of sectors. Theproposed project contributes to the growth of VerticalScope, and to the Canadian economy as a result, byallowing the company to apply cutting-edge research for improving user experience on their forums, and hencethe attractiveness of its service to the businesses.