Fourth Experimental Result on Revues.org corpus level 1
Until now, the dataset of best performance is dataset ver. 2, in the second experiment with 87.86% of overall accuracy.
Based on the know-how that we have acquired from the failures of the previous experiments, we eventually found a specific data setting on which we obtain an enhanced performance in terms of accuracy and also more reasonable labelling result than dataset ver. 2.
The importance of the role of punctuation was verified in the third experiments. We should not ignore it but taking account of all the punctuation types as labels is too much. In our fourth experiment, we propose a compromise that weights more the role of punctuation in learning process than we did on dataset ver. 2.
Our first try with this objective is simply adding two important types of punctuation, COMMA and POINT as features. Since they play actually the major roles in the segmentation of label fields, we first test them as features. So, when a token’s tag is <c>, we check if it is one of comma and point, then if yes, we use it as feature. We take an example of this format (‘/’ is line separator):
Vincent surname / Jean-Marie forename /, COMMA c / 2000 date /, COMMA c / « c /…
With this newly prepared dataset, we have got a small improvement on accuracy (88.3%).
Being encouraged by this result, we tried to extract all the punctuation marks in <c> tag as features. But in this case, the overall accuracy is improved a little (88.5%), but the performance on surname and forename much decreased. Maybe too detailed information negatively influences CRF model learning.
We finally tried to sweep some meaningless labels still existing in learning and test set such as <lb>, <pb>, <hi> and <emph>. They were removed in dataset ver.2 only if there had been the other tags surrounding present token. And that is why there still remain some useless labels. We have replaced them with <nolabel>, which has been already used.
Therefore, our final test here is conducted on the new dataset where <lb>, <pb>, <hi> and <emph> are swept up, then COMMA and POINT are used as unique features. In short, our finally added rule for the preparation of dataset ver. 4 (dataset ver. 3 includes all the variations on the third experiments) is:
“Replace <lb>, <hi>, <pb>, <ptr> and <emph> labels with <nolabel>, then take COMMA and POINT with the label <c> as features of corresponding token”
With this added rule, we got 88.54% on overall accuracy. The following table shows the performance comparison on dataset ver. 1, ver. 2, et ver. 4.