Automatic Semantic Annotation Using Machine Learning

Introduction

Semantic annotation of the web documents is the only way to make the Semantic Web vision a reality. The current Semantic Web meets a bottleneck that there is not much of a Semantic Web due to the lack of annotated web pages. There is such a lack that the Semantic Web is still submerged in the sea of the un-meaningful (un-annotated) web pages.

Semantic annotations are to tag ontology class instance data and map it onto ontology classes. Manual annotation is more easily accomplished today, using authoring tools such as OntoMat (Handschuh, Staab, and Ciravegna, 2002) and SHOE (Heflin, Hendler, and Luke, 2003). However, the use of human annotators is often fraught with errors due to factors such as annotator familiarity with the domain, amount of training, and complex schemas. Manual annotation is also expensive and cannot be used to deal with the large volume of the existing documents on the Web. Automatic semantic annotation is an ideal solution to the problem. However, the fully automatic creation of semantic annotations is also an unsolved problem. Hence, semi-automatic creation of annotations is the method mostly used in current systems.

There are many automatic annotation methods have been proposed, including: A. supervised machine learning based method, B. unsupervised machine learning based method, and C. ontology based method.

• A.

  The supervised machine learning based method consists of two stages: annotation and training. In annotation, we are given a document in either plain text or semi-structured (e.g. emails, web pages, forums, etc.), and the objective is to identify the entities and the semantic relations between the entities. In training, the task is to learn the model(s) that are used in the annotation stage. For learning the models, the input data is often viewed as a sequence of units, for example, a document can be viewed as a sequence of either words or text lines (depending on the specific applications). In the supervised machine learning based method, labeled data for training the model is required.

• B.

  The unsupervised machine learning based method tries to create the annotation without labeled data. For example, Crescenzi, Mecca, and Merialdo (2001) propose a method for automatically generalizing the extraction patterns from the web pages. The generalized patterns can then be used to extract the data from the Web.

• C.

  The ontology based method employs the other knowledge sources like thesaurus, ontology, etc. The basic idea is to first construct a pattern-based ontology, and then use the ontology to extract the needed information from the web page. Some systems also utilize the human general knowledge from common sense ontologies such as Cyc (Lenat and Guha, 1990) and WordNet (Fellbaum, 1998).