Acquiring Annotation Knowledge

There are a variety of techniques which we can use to acquire annotations; the techniques allow a tradeoff between effectiveness of the annotation and the intrusiveness on the user. In all cases, the asset being annotated may be any information segment reachable through the Web: a text file, a multi-media Web page, a procedure, etc.

• User provided annotation: In this mode, the user supplies to HAWK a URL (or other indication of the location of the asset) together with a set of sentences. These sentences are meant to be the answers to a set of questions which the asset answers. HAWK provides the annotations and the information segment to START but monitors the processing to make sure that the sentences can in fact be analyzed correctly. Optionally, the user may provide HAWK a set of questions which she believes the annotations should answer. HAWK will check that each question can be answered and that the asset provided is among the possible answers. Failures to answer the provided questions or failures to parse the annotations initiate interactions aimed at acquiring missing knowledge; notice, however, that this missing knowledge is of a different type, we will discuss acquiring ontologies, domain theories and know-how later.

• Annotations inferred from the content: In this mode, the user provides a useful asset to HAWK without annotation. This is particularly relevant when the ``user'' is in fact a ``Web-crawler''. We will explore several techniques for finding sentences which might act as good annotations, including:
  • Statistical Methods: Statistical techniques are the power behind today's Web indexers. To simplify a complex story, they work by forming a statistical profile of the vocabulary used in a document. Hill climbing techniques can be used to find a set of sentences within the document whose profile best approximates that of the whole document. These sentences might be anticipated to form a good set of annotations.

  • Linguistic Clues: A variety of linguistic markers indicate the presence of topic sentences: a variety of words (e.g ``therefore'', ``consequently'') indicate that a conclusion has been drawn. Such sentences can be anticipated to answer some important questions about the document and therefore might be useful annotations.

• Annotations inferred by the context of use: HAWK will observe the patterns of access of its users. When a link from an annotated page to one without annotations is repeatedly followed, HAWK can ask the user why she followed the link, in particular what question did she seek to answer. The answer to this question could provide a good annotation for the previously unannotated asset.