Don't let's try to break this down : teasing apart lexical chunks

Abstract

Chunk acquisition from the BNC data resulted in chunk lists of hundreds of thousands and occasionally over 1 million chunks. Because of the impracticability of evaluating such a large number of chunks for each measure, evaluation was performed in the following manner: an evaluation text was selected, and all 'good' chunks in that text were determined through a combination of human judgments and comparison with dictionaries, as described later. The chunk databases found by each statistical measure were then used by a program which went through the article and extracted all lexical chunks that occurred in the database. The final list of chunks extracted in this way was then compared to the pre-determined 'good' chunks for each measure, and values of precision, recall, and f-measure were calculated.