Conclusions and further work

The experiment shows that it is possible to build a system that can aid in the drafting lifecycle of amendments using some relatively straight-forward techniques. The system includes a mechanism to validate the correctness of an amendment from a technical perspective. Furthermore, the system aids in the ordering of amendments for voting. Finally, the system provides a method for simulating the effect of an amendment on the actual law.

The main problem is reliable information extraction. Although information extraction algorithms are getting more powerful, there is still noise due to the nature of human language. With XML authoring a hybrid approach can be created where users disambiguate during the authoring process by simply applying markup in case the system gets it wrong. With the addition of XPath to the TokensRegex regular expression language, it is possible to write rules that take manual disambiguation into account.

The described Apache UIMA inspired analysis pipeline architecture works well for this kind of information extraction problems. An observation to make is the described system is essentially a compiler for human language: It takes in the text of an amendment which is essentially lexing and tokenizing. From the tokens, an amendment graph is constructed which is an Abstract Syntax Tree (AST). From the AST an XSLT stylesheet is generated which outputs the modified law. Compiler errors are shown to the user as warnings providing hints that something is not correct just yet.

From a technical perspective, the next step would be to be able to automatically learn patterns. This would make the system scalable across different legal systems without the need of rewriting all the rules from scratch every time. The idea is to augment systems like RAPIER [Califf and Mooney 1997] or WHISK [Soderland 1999] to work with the XML-aware TokensRegex.

So far, we approached this experiment from a technical perspective, although it is based on actual issues encountered in the parliaments today. The next step is to take such a system in production. For a legal system, this would mean writing more rules to deal with edge cases. The biggest hurdle are the legal implications though: some of the features described in this paper may not be used legally in some legal systems due to rules of the parliamentary procedures.