Automatic Assignment of Section Structure to Texts of Dutch Court Judgments

We have successfully demonstrated a method to assign a section hierarchy to documents of Dutch court judgments.

We have described a procedure to assign types to document elements of either title, nr, newline or text block using Conditional Random Fields, reporting an F₁ score of 0.91 and F_0.5 score of 0.91.

We have also reviewed a procedure to organize those elements into a section hierarchy using Probabilistic Context-Free Grammars, reporting an F₁ score of 0.92.

Whether these results are good enough to be used in practice depends on one's tolerance to inaccuracies. As discussed, we rather miss opportunities to enrich data rather than to produce false information, so a low recall is preferable to low precision. The scores obtained for the classifier and parser are promising, but the procedures are not optimized extensively to the corpus, and may be improved to perform within a 5% error margin. In any case, mislabelings do not distort the text in such a way to render it illegible, so we can be somewhat forgiving of errors.