Cohen and Carvalho, 2005
Citation
Carlson, A., S. Schafer. 2008. Bootstrapping Information Extraction from Semi-structured Web Pages. ECML PKDD '08: Proceedings of the 2008 European Conference on Machine Learning and Knowledge Discovery in Databases - Part I, 2008, 195-210, Berlin, Heidelberg.
Online version
Summary
This paper introduces a novel schema for stacked sequential learning. Stacked sequential learning is a meta-learning technique that takes an arbitrary base learner and improves its performance by making it aware of nearby examples. In the experiments they have shown the stacked sequential technique improves performance of both sequential (e.g. conditional random field) and non-sequential algorithms. Their technique can be applied on any base learner and imposes only a constant overhead in training time.
This technique is evaluated on the problem of recognizing the signature section of an email. The dataset contains 617 emails. Each email is represented by a vector of feature vectors. Each element of this vector is a feature vector for each line of the email. A line is labeled as positive if it is part of the email signature, otherwise it is labeled as negative. Hand-crafted features (e.g. "line is blank") are used to represent each line of the email.
The intuition behind their technique is to use global features to infer rules about the local features. For example suppose that we know the name of a set of books. Then by looking at webpages of Amazon.com and by searching the name of books we can infer that the position and font of the book title is the same in most the webpages. We can then use these two features (position and font of book title in web pages) to extract new book titles.
They have described both generative and discriminative approaches for classification and extraction tasks. Global features are govern by parameters that are shared by all data and local features are shared only by a subset of data. For example in information extraction task, all the words in a webpage (without considering formatting) can be considered as global features. On the other hand, features such as position of a text box or color of a text are local features.
They have tested their method on two different datasets. The first dataset contains 1000 HTML documents. Each document is automatically divided into a set of words with similar layout characteristics and then are hand-labeled as containing or not containing a job title. The local and global features for this domain are the same as what we discussed above. The second dataset contain 42,548 web pages from 330 web sites which each web page is hand-labeled as if it is a press release or not press release. The global feature is a set of word in each webpage and local feature is the URL of the webpage. Their experimental result have shown that this approach can obtain high precision and low/moderate recall.
