Third Experimental Result on Revues.org corpus level 1 – Part I
Part I : Attribute analysis
In this and the next postings, we briefly report a number of experiments, which end in failure in terms of accuracy. From these failures, we could successfully move on to the next step of the experiments that have improved performance. The experiments are conducted for the preparation of a reasonable dataset, which includes tag attribute information. Since the tag attributes can be used as one of the labels and features, there would be a lot of strategies to prepare a dataset. To find an optimal strategy, we first try to understand in detail the attributes, their values and roles.
We extract all the types of tag-attribute pairs. Since a tag has sometimes more than an attribute at a time, several types have the form of a tag and multiple attributes. Therefore, a unique type is composed by ‘a tag’, ‘one or more attribute(s)’ and ‘its(their) value(s)’.
For example, the type “title level a” indicates the unique tag <title> which has one attribute ‘level’ having ‘a’ as the value. Another type “title level type m main” is the unique tag <title> that has an attribute ‘level’ with ‘m’ value and the other attribute ‘type’ with ‘main’. In this way, we could extract 188 types of pairs including some spelling errors coming from the manual tagging. We may include these spelling errors in learning procedure to see if small errors can be automatically ignored in modeling then they do not influence on the estimation result.
There are about 160 types of tag-attribute pairs excluding the erroneous pairs. The following table shows the characteristics of some important pairs. “L/F” item indicates if the corresponding attribute can be used as L(label) or F(feature).