Knowledge extraction from webpages

Abstract

This article presents a system to extract Knowledge from webpages by producing semantic annotations. taking into account semantic information from the domain to annotate an element in a webpage implies solving two problems : (1) identifying the syntactic structure of this element in the webpage and (2) identifying the most speci c concept (in terms of subsumption) of the ontology that will be used to annotate this element. Our approach relies on a wrapper-based machine learning algorithm combined with reasoning making use of the formal structure of the ontology.

-

2

Extracting knowledge

Manual annotation

The system is based on machine learning techniques ( g. 1). It learns from some examples the data to be extracted. Here, the examples are webpages annotated by hand according to an ontology. This task is performed in a dedicated environment in which a user is presented a page to annotate and the concepts of the ontology in a browser-like interface. The user then annotates a few occurences of the concepts of the ontology he wants the system to identify. The number of annotations needed depends on the regularity of the page. If the data is strongly structured, like tables, only two or three examples are needed. For example, to extract knowledge about research teams, the SW RC ontology ( g. 2) is loaded along with a page presenting the persons working in a team and the projects they are involved in. The user then annotates some data that are instances of the concept of Person and some instantiating the concept of Project. The output is a marked document in which the annotations are embedded. 2.2

Wrapper induction using the tree structure of the page

The learning algorithm is derived from Kushmerick's work on wrapper induction, which identi es classes of wrappers than can be learnt using a deterministic algorithm with low complexity [ 3 ]. A wrapper is a procedure using the syntactic regularities of a document to extract data. This is particularly suited to semistructured documents like webpages. We have adapted Kushmerick's work to make use of the tree structure of a webpage provided by the w3c's Document Object Model (DOM). This model de nes a path leading to each data in the document. The marked page is taken as input and the similarities between paths leading to data which are instances of the same concept are learnt. The output is a wrapper which is applied to pages in which the tree structure is similar to the example page in order to extract the data and their relationships. In step 3, the relevant wrappers are applied to the documents to extract data. For each extracted data, an instance of the concept it belongs to is added to the KB together with its relationships with other data. The resulting KB is a graph implemented using the Resource Description Framework (RDF) which connects each individual to the ontology and the individuals it is related to. At that point, one problem is that the extracted knowledge is not speci c enough. The reason is that a wrapper must be as generic as possible in order to extract all the relevant data in a document. Therefore, it is induced using the most general concept available. For example, to extract people from a research team according to the SW RC ontology, the wrapper will be designed to recognize any instance of the concept Person instead of more speci c concepts like AcademicSta or Student. The second limitation is that the semantics of the relationships is unknown, since they are extracted using syntactic properties (for instance, two data are related if they have a common parent node). 3

Re ning the knowledge

With an ontology implemented in the Knowledge Representation language OWL, reasoning mechanisms are provided. One of them is instantiation, which given an individual and a set of concepts nds the most speci c concept the individual belongs to, with respect to the subsumption ordering. For example, in the SW RC ontology, two concepts are subsumed by the Person concept and have a role whose ller must be an instance of the Project concept. Therefore, the individuals in the KB that are instances of a Person and have a relationship to a Project must be either instances of AcademicSta or PhdStudent and the relationship is identi ed as being the Project role. The new knowledge is inserted into the KB.

Conclusion

We have presented a system that integrates semantics in the annotation generation process by making use of the reasoning mechanisms provided by the ontology according to which webpages are annotated. This requires not only concept instances but also role instances to be extracted. Initial works on webpage annotation such as Annotea [ 4 ] aimed at enabling collaborative work between people. The need for annotations machines could understand and reason with led to systems, such as S-CREAM [ 5 ], producing semantic annotations according to Description Logics based ontologies. Since manual annotation is a tedious and error-prone task, machine learning system have been proposed; S-CREAM and MnM [ 6 ] implement Amilcare [ 7 ], a semi-automatic tool that produces extraction rules from a corpus to generate concept instances; however, role instances are not dealt with. Recently, fully automatic systems like Amardillo [ 8 ] or CPankow [ 9 ] have been presented that make use of the redundancy of data on the Web. Our system relies on the hypothesis of a of mapping between the syntax and the semantics of a webpage. Since its e ciency depends on the presence of regularities in the structure, pages from research team websites are well suited for this task.

References

1. Kiryakov , A. , Popov , B. , Ognyano , D. , Manov , D. , Kirilov , A. , Goranov , M. : Semantic annotation, indexing, and retrieval . In: International Semantic Web Conference. ( 2003 ) 484 { 499

2. Horrocks , I. , Sattler , U. , Tobies , S. : Practical reasoning for very expressive description logics . CoRR cs.LO/0005013 ( 2000 )

3. Kushmerick , N. , Weld , D.S. , Doorenbos , R.B.: Wrapper induction for information extraction . In: IJCAI (1) . ( 1997 ) 729 { 737

4. Kahan , J. , Koivunen , M.R. : Annotea: an open rdf infrastructure for shared web annotations . In: WWW '01: Proceedings of the 10th international conference on World Wide Web , New York, NY, USA, ACM Press ( 2001 ) 623 { 632

5. Handschuh , S. , Staab , S. , Ciravegna , F.: S-cream-semi-automatic creation of metadata . Proc. of the European Conference on Knowledge Acquisition and Management ( 2002 ) Springer Verlag (submitted version).

6. Vargas-Vera , M. , Motta , E. , Domingue , J. , Lanzoni , M. , Stutt , A. , Ciravegna , F. : Mnm: Ontology driven semi-automatic and automatic support for semantic markup . In: EKAW . ( 2002 ) 379 { 391

7. Ciravegna , F. , Dingli , A. , Wilks , Y. , Petrelli , D. : Adaptive information extraction for document annotation in amilcare . In: SIGIR '02: Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval , New York, NY, USA, ACM Press ( 2002 ) 451 { 451

8. Ciravegna , F. , Chapman , S. , Dingli , A. , Wilks , Y. : Learning to harvest information for the semantic web . In: ESWS . ( 2004 ) 312 { 326

9. Cimiano , P. , Ladwig , G. , Staab , S. : Gimme' the context: context-driven automatic semantic annotation with c-pankow . In: WWW '05: Proceedings of the 14th international conference on World Wide Web , New York, NY, USA, ACM Press ( 2005 ) 332 { 341