We clearly show firstly that refinement improves rules induction, and second, that the use of prior knowledge improves the accuracy of the rules in the case of linguistic structures. This remark may seem trivial, but since, in practice, such knowledge is not used (or so rarely), it may be useful to point it out.
One question that comes up is whether the use of this background knowledge is also useful with other systems or not, especially with systems that perform better than ALLiS [, see][]TSKst00. Does improvements due to the linguistic knowledge only concern ``easy'' structures that ALLiS can not learn and other can? A comparison with other results shows that all the systems have problems with similar structures (noise and coordination). We can point out that some of these systems generally use a combination of several representation schemas, and not only the simple I, O, B categories. We may wonder whether this representation introduces information that can be redundant with the knowledge ALLiS uses.
Except for this refinement mechanism, ALLiS is a basic inductive system with large room for improvements, in particular during the selection of literals (Section 2.3). In CoNLL'00 data, the number of possible literals is small (only POS tag and words) and does not require any sophisticated selection algorithm. In some applications where the number of literals is larger, a selection algorithm would have to be integrated into ALLiS.
Another improvement directly concerns the refinement mechanism. Instead of using a threshold for the mother rule and then learning exceptions, it would be useful to estimate the accuracy of the set {r + exceptions}. This set will be kept if and only if its accuracy is greater than a threshold. Otherwise, the rule r would not be kept, its exceptions being judged too difficult to learn.
This research was funded by the TMR network Learning Computational Grammars (http://www.lcg-www.uia.ac.be/lcg/).