This paper considers the possibility of improving the retrieval of information from legal cases by incorporating semantic elements into the search process. Since legal cases often turn on the cause or causer of an illegal action, it would seem helpful to extract information on the events described within a legal document. To do this, events are extracted using a parser that has been trained on PropBank data (a corpus that is annotated with propositions and their arguments) and on NomBank data (where the nouns are annotated). These corpora were used in the absence of a “sufficiently large, clean corpus of legal texts.” Events are represented in the form REL(arg0, arg1, ..., argN), and “missing arguments are represented by an asterisk wildcard”--for example, “reply(DEFENDANT, *, enjoy, immunity).”

Event extraction was performed on two Canadian Supreme Court cases. One of the authors worked with a legal professional to prepare a baseline for comparison. The following criteria were used to judge the retrieval of events: accuracy, capture of legally or factually important aspects, and completeness of meaning captured from each sentence. Legal professionals rated the results produced by the system. By filtering the extracted events--for example, by removing events for which there was no agent specified--the ratings showed a 74 percent accuracy rate, 72 percent legal/factual importance, and 86 percent for completeness.

These encouraging results suggest that extracting propositional information that is appropriate to the domain of interest does improve retrieval.