Product Name Recognition In Online Advertising Delivery

With the rapid growth of the Internet, more and more sellers use the Internet to deliver promotional marketing messages to consumers. Consequently, a new business model-Online Advertising-has emerged. Compared with classic advertising, it has the advantages of low cost, interaction, quantitative analysis, etc. Because of these characteristics, online advertising has developed quickly and received keen attention in research community.How to select the most suitable one from many advertisements according to the context of user is an important problem when advertising agencies deliver advertisements. To solve this problem, product names which appear in advertisement and the context of user should be paid much attention. The product name in advertisement tells us which product this advertisement is for and the product name in the context of a user gives a strong indication about what the user is interested in. These two types of data are useful for solving the problem. Besides this, in order to avoid producing advertisements for prohibited items, the product name in advertisement should be heeded. In this dissertation, some key issues related to product name recognition in online advertising delivery are investigated deeply. The main contents of this dissertation are as follows:1. One product can be named in several ways in Chinese. Although both names specify the same product, the one used in the context of user may be different with the one in advertisement. If they cannot be recognized as the names of same product, this advertisement will be missed when advertising agency selects advertisements for this user. In this condition, the profit of advertising agency is influenced. To recognize the aliases of product, we developed a system for collecting different Chinese names of same product automatically. Someone provides two Chinese aliases of a product as seeds, and then our system can find more aliases of the product from the Internet. The system is based on a special phenomenon on shopping web pages discovered by us. Using this phenomenon, different candidate names of same product can be extracted. To remove irrelevant words from these candidates, we proposed a filtering method based on set operations and a modifier removing method based on pointwise mutual information. We also abstracted the relations among candidates to a graph, and used random walk with restart, Simrank++to calculate the reliability of candidates on it.2. To filter out the advertisements sold prohibited items, advertising agency keeps list of the different Chinese names of each prohibited item. In this way, the aliases of prohibited items can be recognized. We developed a system for collecting different Chinese names of same prohibited item automatically. Someone provides two Chinese aliases of a prohibited item as seeds, and then our system can find more aliases of the prohibited item from Internet. Because the number of web pages sold prohibited items is often smaller than the number of web pages sold products, the reliability calculation method for product names is not feasible for prohibited items. In this condition, we proposed a reliability calculation method for prohibited items based on Google distance.3. In order to utilize product name in the context of user for online advertising delivery, the name of product should be recognized in the context first. Supervised labeling method requires a large amount of manually annotated training data. The quantity of public training data which can be used for product name recognition is small, which makes product name recognition difficult for supervised method. The features conveying rich semantic information can help to improve the performance of supervised product name recognition. We explored these features such as clustering-based word representations and distributed word representations when recognizing product names with a small amount of training data.