Mining the Web for Relations between Digital Devices using a Probabilistic Maximum Margin Model

Mining the Web for Relations between Digital Devices using a Probabilistic Maximum Margin Model Oksana Yakhnenko Iowa State University Ames, IA, 50010 oksayakh@cs.iastate.edu Barbara Rosario Intel Research Santa Clara, CA, 95054 barbara.rosario@intel.com Abstract Searching and reading the Web is one of the principal methods used to seek out infor- mation to resolve problems about technol- ogy in general and digital devices in partic- ular. This paper addresses the problem of text mining in the digital devices domain. In particular, we address the task of detecting semantic relations between digital devices in the text of Web pages. We use a Na¨ıve Bayes model trained to maximize the margin and compare its performance with several other comparable methods. We construct a novel dataset which consists of segments of text extracted from the Web, where each segment contains pairs of devices. We also propose a novel, inexpensive and very effective way of getting people to label text data using a Web service, the Mechanical Turk. Our re- sults show that the maximum margin model consistently outperforms the other methods. 1 Introduction In the digital home domain, home networks are moving beyond the common infrastructure of routers and wireless access points to include application-oriented devices like network attached storage, Internet telephones (VOIP), digital video recorders (e.g., Tivo), media players, entertainment PCs, home automation, and networked photo print- ers. There is an ongoing challenge associated with domestic network design, technology education, de- vice setup, repair, and tuning. In this digital home setting, searching the Web is one of the principle methods used to seek out information and to resolve problems about technology in general and about dig- ital devices in particular (Bly et al., 2006). This paper addresses the problem of automatic text mining in the digital networks domain. Under- standing the relations between entities in natural lan- guage sentences is a crucial step toward the goal of text mining. We address the task of identifying and extracting the sentences from Web pages which ex- pressed a relation between two given digital devices in contrast to sentences in which these devices co- occur. As an example, consider a user who is looking for information on digital video recorders (DVR), in particular, on how she can use a DVR with a PC. This user will not be satisfied with finding Web pages that simply mention these devices (such as the many products catalogs or shopping sites), but rather, the user is interested in retrieving and read- ing only the Web pages in which a specific relation between the two devices is expressed. The user is interested to learn that, for example, “Any modern Windows PC can be used for DVR duty” or that it is possible to transfer data from a DVR to a PC (“You can simply take out the HD from the DVR, hook it up to the PC, and copy the videos over to the PC”).1 The specific task addressed in this paper is the fol- lowing: given a pair of devices, search …

Content truncated to 500 words