Information Technology Reference
In-Depth Information
Design Implications
The results of this evaluation show that handwritten annotations of lecture slides
cannot be recognized with a high accuracy using a standard handwriting recognition
engine.
4
Due to this low recognition accuracy, the interaction design of CoScribe
does not require handwriting recognition. In particular, handwriting is not displayed
as text but as the original handwriting. Nevertheless, handwriting recognition is used
in the background. The text of each newly created or modified annotation is auto-
matically recognized.
We have shown that the recognition performance for domain-specific terms can
be significantly improved by using domain-specific dictionaries, which are automat-
ically created from the annotated document. The recognized text can then be used
for full-text search within the annotated document.
4
Recently, Liwicki et al. [80] presented a similar approach for improving the recognition of hand-
written annotations on paper documents, using the same handwriting recognition engine. Instead
of text contained within the document, they extracted text from a personal knowledge base to im-
prove the vocabulary of the recognizer. Their approach could reduce the overall word error rate by
4 %. They report a better baseline performance (approx. 30 % word error rate) than found in our
results. We assume that this is due to the fact that their annotations contain only text; in contrast in
our experiment many annotations contained sketches.
